Caring for Vincent A Chatbot for Self-compassionMinha Lee Sander Ackermans Nena van As Hanwen Chang Enzo Lucas and Wijnand IJsselsteijn

Human-Technology InteractionEindhoven University of Technology

Eindhoven Noord Brabant the NetherlandsMLeeWAIJsselsteijntuenl

NLVAsEMLucasSCAAckermansHChangstudenttuenl

ABSTRACTThe digitization of mental health care holds promises of af-fordable and ubiquitously available treatment eg with con-versational agents (chatbots) While technology can guidepeople to care for themselves we examined how people cancare for another being as away to care for themselvesWe cre-ated a self-compassion chatbot (Vincent) and compared be-tween caregiving and care-receiving conditions Care-givingVincent asked participants to partake in self-compassion ex-ercises Care-receiving Vincent shared its foibles eg embar-rassingly arriving late at an IP address and sought out adviceWhile self-compassion increased for both conditions onlythose with care-receiving Vincent significantly improvedIn tandem we offer qualitative data on how participantsinteracted with Vincent Our exploratory research showsthat when a person cares for a chatbot the personrsquos self-compassion can be enhanced We further reflect on designimplications for strengthening mental health with chatbots

CCS CONCEPTSbullHuman-centered computingrarrHCI theory conceptsand models Empirical studies in HCI

KEYWORDSChatbot compassion mental health well-being positivecomputingACM Reference FormatMinha Lee Sander Ackermans Nena van As Hanwen ChangEnzo Lucas and Wijnand IJsselsteijn 2019 Caring for VincentA Chatbot for Self-compassion In CHI Conference on Human Fac-tors in Computing Systems Proceedings (CHI 2019) May 4ndash9 2019

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies are notmade or distributed for profit or commercial advantage and that copies bearthis notice and the full citation on the first page Copyrights for componentsof this work owned by others than ACMmust be honored Abstracting withcredit is permitted To copy otherwise or republish to post on servers or toredistribute to lists requires prior specific permission andor a fee Requestpermissions from permissionsacmorgCHI 2019 May 4ndash9 2019 Glasgow Scotland UKcopy 2019 Association for Computing MachineryACM ISBN 978-1-4503-5970-21905 $1500httpsdoiorg10114532906053300932

Glasgow Scotland UK ACM New York NY USA 13 pages httpsdoiorg10114532906053300932

1 INTRODUCTIONApproximately 1 out of 10 people need psychiatric careworldwide yet only 70 mental health professionals are avail-able for every 100000 people in high-income nations andthis number can drop to 2 for every 100000 in low-incomecountries [40] Psychiatric illnesses such as depression areof a growing concern for many societies yet adequate carefor those in need is often not sufficiently provided [39] Thustechnology offers promising means for treating mental ill-nesses like depression for example through mobile apps[1 21] chatbots [15] or virtual reality [14] However thesetechnological solutions thus far do not adequately cover twoaspects (1) they often target what users can do for them-selves and what is missing is what users can do for anotherbeing as a potential treatment for themselves and (2) theydo not address pre-emptive care for strengthening mentalhealth without necessarily assuming diagnosed disordersthat people may or may not associate with Usually the fo-cus is on what should be ldquofixedrdquo eg depressive symptomsand the target is the person with these symptoms We re-versed this framework with a chatbot named Vincent thatpeople could care for and be cared by agrave la TamagotchiWe pose the question ldquowhen bots have psychological is-

sues can humans care for them and if so howrdquo By doingso we offer exploratory results on (1) how caring for a chat-bot can help people more so than being cared by a chatbotand (2) how aiming for an increase in self-compassion canpotentially strengthen psychological well-being which is aholistic preventative way of envisioning mental health carePeople can feel psychologically vulnerable in varying waysand to varying degrees in everyday life whether or not theychoose to use clinical terms to label how they are or feelMental health care can be geared towards prevention ratherthan treatment by fortifying peoplersquos resilience to psycho-logical ill-being Self-compassion is especially suitable forpre-emptive care because it is causally linked to well-being[52]We explored if two weeks of human-chatbot interaction

would result in greater self-compassion for our participants

a non-clinical sample As the sample choice indicates ourfocus was not on clinically defined symptoms of mental ill-being greater self-compassion can benefit people in generalnot just those with mental health disorders We comparedbetween caregiving (CG) and care-receiving (CR) Vincents astwo conditions Our research question thus follows Are thereself-reported differences in self-compassion states after inter-acting with a CG chatbot and a CR chatbot for a non-clinicalsample What implications do these patterns of interactionsuggest We hypothesized that both CG and CR conditionswould increase participantsrsquo self-compassion and aimed forquantitative and qualitative analyses

2 BACKGROUNDWe start with related works on caregiving (CG) and care-receiving (CR) robots and then we touch on how this cantranslate to chatbots After that we define compassion andself-compassion in light of positive computing (technologyfor well-being) We cover that self-compassion can bringabout well-being and posit that chatbots can be vehicles forimproving peoplersquos self-compassionComputers are social actors (CASA) even when people

know they are interacting with machines they tend to treatmachines in a social manner [31] People reciprocate helpwhen a computer was helpful to them before [16] and at-tribute personality traits to computers that communicatewith them only via text [28] The CASA paradigm is a help-ful albeit broader framework for understanding caregiving(CG) and care-receiving (CR) behaviors of machines

The comparison between CG andCR chatbots has not beenpreviously explored but there are related works in human-robot interaction (HRI) People tend to care for a robot byassuming and anticipating its needs [11] In the context ofldquolearning by teachingrdquo ie when students learn the materialby teaching about it a CR robot that acted as childrenrsquosstudent was effective in helping students retain knowledge[49] In a study with an older population a robot that askedfor help from humans was accepted as a mutual companionrobots that people could care for and be cared by may pavenew grounds for assistive technologies that aim for reciprocalcare between humans and robots [26] There are emotionalpsychological and physiological costs and benefits in caringfor another being eg comfort one gets from a pet vs costsof caring for a pet Yet human investments may not havesuch equitable pay-offs in HRI which is a caveat that requiresfurther research [11]As with robots chatbots have the capacity to give and

receive care They do not have the same level of physicalpresence as robots but uni-modal (text or voice) interactionscan still be behaviorally powerful while being less costly todesign and deploy An added benefit of chatbots is that theyexist on messaging platforms like Facebook Messenger or

Slack that many people already use [24] which translates tohigher accessibility to chatbots compared to robots An earlyexample of a chatbot ELIZA acted as a therapist and somepeople believed that they were interacting with a humanbased on simple text-based chats [50] Nowadays chatbots(both voice and text-based) are re-emerging as interactiveentities that serve as all-in-one assistants like Applersquos Siri oract as specialists in specific contexts eg helping users shopfor groceries [10] or for therapeuticself-help purposes [38]A recent example of a chatbot for mental health care is

Woebot It was designed to help combat depression [15]1After two weeks of interaction Woebot reduced signs ofdepression for young adults who self-reportedly suffer fromdepression or anxiety (p = 001) while the control group thatwas given an ebook called ldquoDepression in College Studentsrdquodid not show improvements [15] Woebot gave care to partic-ipants but did not receive care from its interactants In onestudy that looked into self-compassion clinically depressedparticipants were told to be compassionate towards a virtualagent that they later embodied to receive compassion fromthemselves meaning participants gave and received carefor and from themselves through VR [14] This led to low-ered self-criticism lowered symptoms of depression and in-creased self-compassion [14] To conceptually combine thesestudies [13 15] a chatbot that gives care andor receives carefrom people could potentially increase self-compassion

Compassion is a moral emotion [19] or motivation to freeourselves and others of suffering with loving-kindness [18]by having concern [37] or a caregiving approach [6] towardsliving beings It is at the heart of Mahayana Buddhism asexpressed through stories in key texts like the Lotus Su-tra [27 42] Schopenhauer influenced by Buddhism extolledcompassion as the basis of morality and found it celebrated inmany cultures eg ldquoat Athens there was an altar to Compas-sion in the Agora [] Phocion (ancient Athenian politician)[] describes Compassion as the most sacred thing in humanliferdquo [46 p 98-99] Compassion and empathy are associatedbut are not the same Empathy allows people to relate tootherrsquos suffering cognitively and affectively [23] Howeverempathic concern for others can lead to empathic distress astate of over-identifying with sufferers that leads to vicariouspain without prosocial altruism to help [6 37] Compassionbuilds on such empathic connections when one can relate tosufferers in a healthy way without empathic distress [6 37]

Self-compassion is practiced by being kind to oneself witha balanced awareness of onersquos feelings and recognizing thatone is interconnected with others [32] There are three sup-porting elements Self-kindness over self-judgment is to havea forgiving attitude towards onersquos own faults and to embrace

1Woebot - httpswoebotio

onersquos suffering with understanding connectedness over isola-tion is to view onersquos life as intertwined with other lives ratherthan to see onersquos experiences as unrelated or irrelevant togreater humanity mindfulness over over-identification is tobe aware of onersquos negative emotions in a balanced mannerthan to excessively identify with them [52]Onersquos gender may influence self-compassion A meta-

analysis concluded that women score lower than men onself-compassion and the gender difference was especiallypronounced when sampled studies had more ethnic minori-ties [51] Women reportedly have greater empathy than men[23] and they are more likely to be more self-critical thanmen [51] To add women who provide empathy as socialsupport can feel drained or distressed [22] Yet a study witholder adults demonstrated that older women have greatercompassion than older men [29] While peoplersquos gender ageand possibly ethnic minority status may impact their self-compassion practicing self-kindness connectedness andmindfulness can help individuals be more compassionatetowards themselves and others

In clinical settings people who experiencemental ill-beingcan benefit from self-compassion [17] Self-compassion isalso strongly connected to well-being for the general popu-lation [52] Well-being refers to mostly feeling more positiveaffect than negative affect and being satisfied with onersquos lifefactors like income gender age or culture influence onersquoswell-being only minutely [30] Thus having a good balancebetween onersquos psychological social and physical capabili-ties to deal with lifersquos difficulties is important for well-beingrather than having a static life without suffering [12] (noris this realistic) Through an awareness of onersquos and othersrsquosuffering without being overwhelmed by empathic distresscompassion is developed [18 47] Caring for or being com-passionate towards others has been shown to increase onersquosown self-compassion [5] Yet could the same effect be foundwhen people care for technological entities Technology canpotentially be a means to achieve self-compassion and byextension well-beingWe return to the question ldquowhen bots have psychologi-

cal issues can humans care for them and if so howrdquo toemphasize that (1) anthropomorphic realism of artificial en-tities is not required for humans to develop a caretakingstance towards them and (2) machinesrsquo mimicry of peoplersquospsychological ill-being can help ascertain why certain psy-chological traits are labeled as issues When we observe howpeople take care of unwell chatbots we may uncover howthey themselves want to be treated Machines therefore donot need to pass the Turing test for the purpose of positivecomputing or technology for well-being [6]

In order to uncover peoplersquos psychological responses tochatbots particularly in relation to modulating peoplersquos self-compassion we asked participants to interact with a chat-bot Vincent designed to engage in care-giving versus care-receiving conversations Exploring simulated psychologicalstates via technological entities like Vincent is a way to en-vision positive computing In addition our approach focuseson pre-emptive mental health care We now turn to howwe designed our study how we built the two Vincents andpresent our results

3 METHODWe utilized quantitative and qualitative methods to best un-derstand our data We compared self-compassion scores be-fore and after two weeks of interaction and examined if theCR and CG conditions showed any difference Our mixedlongitudinal design was supplemented by thematic [3] andinterpretive analyses [48]Our qualitative analysis was performed on participantsrsquo

open-ended responses to Vincentrsquos questions eg ldquocan youremember something that went well for you recentlyrdquo (CGVincent) ldquocan you think of something that will cheer me upa bitrdquo (CR Vincent) and on post-experiment open endedresponses about the experience in general We coded deduc-tively on the sub-scales of the self-compassion scale and weallowed for inductive themes to emerge [3 48] Four codersanalyzed the data and a fifth annotator broadly checkedfor coherence or incoherence resulting in a structured it-erative process Our quantitative and qualitative measure-ments therefore corresponded with each other to triangulatevarying insights of the same phenomenonmdash self-compassionthrough human-chatbot interaction

Participants and groupsWe conducted a power analysis to estimate our sample sizeOur effect size is based on an aforementioned study [14] thatmeasured self-compassion of a clinical population (N = 15)in an embodied VR experiment with the partial eta-squaredof 036 at p = 002 [14] which gave us a sample size of 68with a power of 90 and an error probability rate of 005 Weplanned for t-tests and thus the transformed effect size viaeta-squared to Cohenrsquos d was 1487 which was reduced toa more realistic 08 We had 67 participants (F = 29 M = 36undisclosed = 2) with the mean age at 251 years (SD = 57range = 19 - 48)We recruited people through the participantsdatabase of the Eindhoven University of Technology Allparticipants completed the self-compassion questionnairebefore they interacted with Vincent Then they were dividedto two conditions so that the average self-compassion scoreswere evenly distributed at the start From the lowest scoringto the highest scoring participants we divided all into eitherthe CR or the CG condition in an alternating manner This

practice resulted in a relatively even gender distribution forboth conditions (CR M = 18 F = 14 undisclosed = 1 CG M= 18 F = 15 undisclosed = 1)

Chatbot implementation

Figure 1 Care-receiving Vincent

Vincent2 was built with Googlersquos Dialog flow that wasintegrated to Facebook Messenger3 We purposefully did notvisually design Vincent (the profile icon showed a ldquoVrdquo) todrive Vincentrsquos personality by what was said rather thanhow Vincent looked Participantsrsquo responses did not changeVincentrsquos reactions The usage of limited pre-set responses(Figure 1) was to continue Vincentrsquos narrative whilst allow-ing participants a choice between relevant responses (Woe-bot also interacted with its users in a similar fashion [15])This allowed for greater comparability between participantsin each condition

We had eight scenarios each for CG or CR Vincent and 6neutral scenarios (total 22 scenarios) The neutral scenarioswere the same for both Vincents and aimed for adherenceto daily touch-points in an entertaining way For exampleone neutral scenario is about world records ldquo[] I knew you

2Vincentrsquos Facebook page - httpswwwfacebookcomvincentthebot3Dialog Flow Facebook Integration - httpsdialogflowcomdocsintegrationsfacebook

were an interesting bunch of carbon-based beings but appar-ently you have this thing called lsquoworld recordsrsquo [] [] whatwould you like to be in the world records book forrdquo BothVincents used images and emojis because visual icons arewidespread in digital messaging to express emotions [43]Weused appropriate punctuation marks and positively and neg-atively valenced syntax in accordance to previous researchon text-based emotion expression and detection [20] Mostinputs were close-ended (Figure 1) but we allowed open-ended free inputs at least once per interaction We aimedfor a meaningful comparison in two ways In each condi-tion participantsrsquo interactions were designed to be the samewith limited set of possible reactions to Vincent Betweenthe two Vincents we wanted to clearly distinguish betweenthe recipient and the giver of care Below are excerpts fromCG and CR Vincents (all scenarios are in the supplementarymaterial)

CG Vincent [] see if you can think of a kinder more caringway to motivate yourself to make a change if needed Whatis the most supportive message you can think of thatrsquos in linewith your underlying wish to be healthy and happy Try towrite it belowUser [free input]

CR Vincent What do you think am I the dumbest bot yoursquoveever seen or whatU [free input]V Am I being too hard on myselfU [free input]

CG Vincent was modeled after Woebot [15] supplementedby self-compassion exercises [34] CR Vincent was based onthe Self-Compassion and Self-Criticism scale [13 14] Thescale has scenarios like job rejections unpaid bill remindersand being late to a meeting which were converted to fita chatbot (Figure 1) The eight items of the scale were in-terweaved for conversational storytelling By doing so wejuxtaposed issues that students (the majority of our sample)can face eg distress about failing an exam that Vincentunderwent CR Vincent narrated its story over time by ad-mitting its mistakes feeling inadequate compared to otherchatbots confessing a former lie and asking for confiden-tiality in sharing a non-trivial worryI got a reminder from the server that hosts me Itrsquos like my

house so to say [] I forgot to pay the server fee on time[][] It wouldrsquove taken me only 0004 seconds to make the trans-action you know since Irsquom a robot and all [] this never seemsto happen to the other chatbots on my serverVincent brought up this scenario again later with new

informationRemember our talk a couple of days ago About me forgetting

to pay my server fee in time [] I kind of lied to you [] Ididnrsquot tell you this before because I was a little embarrassedabout it Can you promise me that this stays between us []Irsquove been applying for different jobs and just today I receivedmy third rejection email already The reason that I couldnrsquotpay my bill was because Irsquom running out of money And if Idonrsquot find a job soon Irsquoll get kicked off my server

In sum CG Vincent guided participants through activitieswhile CR Vincent opened up about its everyday mistakesor worries to get support from participants CG Vincentsought to trigger self-compassion in users themselves and CRVincent gave participants opportunities to be compassionatetowards a chatbot

MeasurementsWe collected basic demographic information ie genderage and previous experience with a chatbot at the startOur main measurement was the self-compassion scale on afive point scale [32] with six sub-components (self-kindnessvs self-judgment common humanity vs isolation mind-fulness vs over-identification) which was deployed twicebefore and after the experiment We added a measurementfor opinions about the agent on a seven point scale [4] todetect irregularities between how CG and CR Vincents maybe perceived Our final scale was the Inclusion of Other inthe Self (IOS) Scale a single item on a seven point scale [2]to check howmuch participants identified with Vincent post-hoc We also kept track of two additional aspects to gaugeengagement quality One is the error rate ie the number oftimes Dialogflow crashed during the interaction sometimesrequiring an experimenter to momentarily ldquowizardrdquo [9] torestart the interaction4 The other is the total word countper participant on open ended answers

ProcedureWe built Vincent and wrote our initial scenarios to be testedOur pilot study of three days was with voluntary participants(N = 12) personally recruited by experimenters We checkedif scenario categories (caregiving care-receiving and neu-tral) were clear by asking participant to guess Vincentrsquosintentions and goals per scenario Based on this we onlyadapted neutral scenarios Then we recruited participantsfor the actual experiment Our email invitation was sent outto the TU Eindhoven participant database and interestedparticipants joined the experiment if they used Facebookand could freely express themselves in English The emailcontained a link to Vincentrsquos Facebook page that partici-pants had to go to for a guided tutorial on the experimentpayment information and the informed consent form This

4Restarts happened 37 out of 938 interactions (14 days 67 participants) or394 of the time

form noted that experimenters will look at participantsrsquo dataand that third party technology providers Vincent relied onie Facebook and Google have access to the informationWe added that participantsrsquo personally identifiable informa-tion will not be shared for publication purposes and thattheir voluntary participation means that they can drop outof the study at any point We also stated that Vincent is nota therapist and that they should seek professional help if anypsychological issues are experienced though we targeted anon-clinical population

After the guided tutorial participants filled in the first setof questions on demographic information and self-compassionThen they were assigned to either CG or CR Vincent Theyall began the experiment on the same day For two weeksVincent greeted and sent a password daily and participantshad to repeat the password to Vincent to start the daily inter-action eg Hey me again ) Tell me tHup to start our littletalk After two weeks participants filled in the final surveyon self-compassion the IOS scale opinion on the chatbotdetails for compensation as well as additional comments orfeedback they were willing to share Participants were thenpaid through bank transfer

4 RESULTSWe first present our quantitative analysis and then move onto how people talked with Vincent the qualitative angle Allrelevant tables and graphs as well as CG and CR Vincentsrsquoscenarios are in the supplementary material

Quantitative analysisBefore we forged ahead with hypotheses testing we lookedinto engagement levels of all participants to detect outliersWe had three outliers participants who had less than 20minutes of total interaction time with Vincent Only readingwhat Vincent sent not including giving a response shouldtake one to two minutes per day We expected a minimum of20 to 28 minutes of interaction for two weeks (for all partici-pants the average total time was 36 minutes SD = 10) Ouroutliers spent in total 15 18 and 19 minutes each the threelowest total interaction time Correspondingly their totalword count to open responses reflected low engagement at27 22 and 27 words for the total duration of the experimentthe three lowest total word count out of all participants

We conducted two one-tailed dependent t-tests [7] to an-swer our hypotheses (we set p at 005 with the confidenceinterval of 95) CG Vincent did not result in significantchange (t(31) = -0572 p = 0286 Cohenrsquos d = 007) whencomparing before (M = 3135 SD = 0630) and after (M =3180 SD = 0628) the two weeks but the direction detectedis positive (normality assumed for prior (W = 0987 p = 0978)and post (W = 0989 p = 0982) scores) CR Vincent did showa significant difference (t(31)= -197 p = 0029 Cohenrsquos d

= 02) between prior (M = 3137 SD = 0613) and post (M= 3257 SD = 0558) scores for self-compassion (normalityassumed for prior (W = 0979 p = 0764) and post (W = 0946p = 0114) scores)We conducted exploratory analyses to better understand

our data Through a repeated measures ANOVA we checkedfor the effect of time prior and post self-compassion scores(F(1 62) = 2768 p = 0101 η2 = 0043) Then we checked forthe effect of condition ie CG or CR (F(1 62) = 075 p = 0785η2 = 0001) and the interaction between time and condition(F(1 62) = 0580 p = 0449 η2 = 0009) None were significantWe additionally checked for time condition timeconditioneffects on the three components of self-compassion self-kindness common humanity and mindfulness Only theeffect of time on common humanity was significant F(1 62)= 6059 p = 0017) The table for all statistical tests is in thesupplementary materialWe further dissected our data by gender because previ-

ous research showed that women may score lower on self-compassion [51] Indeed female participants had lower self-compassion scores (M = 305 SD = 013) than men (M = 326SD = 009) at the start but not significantly so (t (4935) =113 p = 026 d = 029) according an independent unequalvariance t-test (normality assumed) We then compared postand prior scores for men and women Menrsquos self compassionscores increased only by 002 as a difference in means andshowed no significant increase (t (33) = -025 p = 040 d =004) However womenrsquos scores showed a significant differ-ence (t (27)= -206 p = 002) between 305 (SD = 071) as thestarting score and 319 (SD = 065) as a posterior score forself-compassion When we scrutinized the gender differencebetween CG and CR Vincents we noticed a more dramaticdifference Women with CR Vincent showed a highly signifi-cant change (t(13) = -289 p = 0006 d = 077) compared towomen with CG Vincent (t(13) = -033 p = 037 d = 009)and both were normally distributed

We wanted to check if mainly gender was at stake or if itwas simply a difference between low vs high scorers on priorself-compassion levels We thus divided all participants intotwo groups based on the average self-compassion score at thestart 314 Those who scored above this were high scorers(M = 371 SD = 037 N = 28) those who scored below werelow scorers (M = 269 SD = 032 N = 36) We included oneparticipant with the average score in the high-scoring groupand this had no impact on significance reached Low scorersgreatly increased their self-compassion scores in terms ofsignificance (t(35) = -341 p = 00008 d = 057) but highscorers did not show improvements (t(27) = 110 p = 086 d= 018) Yet normality was not assumed for both low-scorers(W = 093 p = 003) and high-scorers (W = 091 p = 002)since we divided a normally distributed group into two Thuswe performed a two-sample Wilcoxon rank-sum test to see

if there was a significant difference between low scorers andhigh scorers which was the case at z = 286 and p = 0004 CRVincent improved low scorersrsquo self-compassion significantly(t(17) = -320 p = 0003 d = 075) compared to a marginalsignificance for CG Vincent (t(17) = -175 p = 005 d = 041)There were normal distributions observed for prior and postself-compassion levels of low and high scorersA potential explanation for why low scorers improved

more than high-scorers is regression to the mean Howeverin published literature the average self-compassion score isbetween 25 and 35 [33] and our low scorers have a prior av-erage self-compassion of 269 If regression to the mean is anexplanation we would also expect high-scorers to end witha lower mean yet this is not the case Our high-scorers hadan average prior score of 371 (above average [33]) and theirscores did not decrease after the experiment This may be aceiling effect The low-scorersrsquo improvement is still there itis a highly significant effect even with the Bonferroni correc-tion for all tests (p = 00008) with the post-hoc power of 097The data supports that Vincent enhanced self-compassionfor low-scorersWe had two additional scales one to check if people per-

ceived CG and CR Vincent in a relatively similar way [4]and the other to see how much participants identified withVincent [2] The survey on participantsrsquo opinion of the agentincluded caring likability trustworthiness intelligence dom-inance and submissiveness as items [4] about Vincent BothCG and CR Vincents were perceived to be fairly analogoussave for dominance (α = 048 CG Vincent M = 2677 SD =0794 CR Vincent M = 2656 SD = 0700) and submissiveness(α = 024 CG Vincent M = 2448 SD = 0559 CR Vincent M= 2396 SD = 0636) though none showed a significant dif-ference between two conditions The the Inclusion of Otherin the Self (IOS) scale [2] displayed that participants moreclosely related to CR Vincent (M = 348 SD = 148) thanCG Vincent (M = 306 SD = 139) but not significantly so (t= -115 p = 025 d = 029) Both conditions were normallydistributed (CG W = 096 p = 034 CR W = 097 p = 064)

To summarize our hypothesis that CG Vincent increasesself-compassion was not supported (p = 0286) but the hy-pothesis that CR Vincent increases self-compassion was sup-ported (p = 0029) Our exploratory analyses captured threeunderlying influences on this finding First our ANOVA testsrevealed that the only significant aspect was time as an inde-pendent variable affecting common humanity one elementof self-compassion (p = 0017) Second gender may be a con-tributing factor with women demonstrating a significantincrease in self-compassion (p = 002) for both conditionscombined but not men (p = 040) To add only CR Vincentdemonstrated a highly significant change for women (p =0006) unlike women who interacted with CG Vincent (p =037) Third regardless of gender those who started out with

a low self-compassion score exhibited the most significantchange (p = 00008) for both conditions together Low-scorersmore significantly improvedwith CR Vincent (p = 0003) thanwith CG Vincent (p = 005)

Put together CR Vincent more effectively increased self-compassion than CG Vincent most likely through a signif-icant change in participantsrsquo sense of common humanitymore so than self-kindness and mindfulness (as our ANOVAtest showed) Finding common humanity can be inclusive ofchatbots Women specifically those with CR Vincent weresignificantly more affected than men However low-scorersof both genders benefited the most compared to high-scorersespecially those with CR Vincent CG and CR Vincents werenot perceived to be significantly different except for a lowersimilarity regarding the dominance-submissive trait Par-ticipants in the CR condition may have felt that Vincent ismore like them (ISO scale) though this difference was notsignificant

Qualitative analysisFor our qualitative analysis we used the corpus of free re-sponses that participants had typed during their interactionswith Vincent We will first present a descriptive analysisof the interactions that people had with CG and CR Vin-cents followed by our thematic analysis [5] and interpretiveanalysis [48] Participantsrsquo responses to CR Vincent were onaverage 9353 words (SD = 4796) and to CG Vincent 11247words (SD = 6399) for two weeks While our data set is notabundant in terms of word count we believe a qualitativelook at how participants interacted with Vincent is valuable

Descriptive analysis CGVincent guided participants throughself-compassion exercises [34] that did not require them toactively voice aspects of self-compassion they mainly hadto read about it This resulted in fewer instances of self-compassion themes in CG Vincent since they only occurredwhen Vincent asked participants to comfort themselves Toadd participantsrsquo willingness to engage with CG Vincentrsquosprobes differed It asked specific questions or provided ashort task for free input eg ldquowrite down a difficulty youhaverdquo Many answers to this were short ldquowake up earlyrdquoor ldquoI often feel alonerdquo Some participants opened up moreldquoI have a difficulty in expressing myself when I am underdifficult situationsrdquo or ldquoI am studying abroad far from homeand family and friends Different culture language edu-cational standardrdquo CG Vincent asked a follow-up questionldquohow did it make you feelrdquo we again got simple answerslike ldquogoodrdquo or ldquonormalrdquo or longer expressions ldquoI feel morerefreshedrdquo or ldquonot really better or worserdquo In other instancesCG Vincent allowed participants to give themselves simpleself-assurance ldquoI can do itrdquo plan for long-term goals ldquoonceI have graduated I can schedule a good routine to target a

fitter and healthier lifestyle Just hang in there a little longerrdquoor dig deeper ldquopeople around me would be happier if I washappierrdquo Thus CG Vincent provided a daily touch point forself-reflection admittance of everyday suffering ldquopep-talkrdquoor deeper self-insight which may or may not directly relateto self-compassion for all participants

In contrast CR Vincent frequently asked for help and con-sequently received many self-compassion related answersThe narrative was the focus for CR Vincent It was able tobecome more vulnerable over time by admitting its own ev-eryday hardships as a chatbot which led it to seek opinion oradvice For example CR Vincent asked ldquowhat do you thinkam I the dumbest bot yoursquove ever seen or what Am I beingtoo hard on myselfrdquo To this participants responded in dif-ferent ways ldquoI think yoursquore the funniest bot that Irsquove everseen mdash yes you are in some situationsrdquo ldquoNo but you shouldexpect (that) a bot is very smartly programmed and know allmdash Maybe I do not knowrdquo or ldquothe world needs bots like youAnd itrsquos usual to get rejected sometimes just keep on goingand yoursquoll find a job soon enoughrdquo However CR Vincentrsquoscries for help did not always result in necessarily compassion-ate replies Many users stuck to pragmatic answers relatedto the topic of the problem Even though all of CR Vincentrsquosscenarios were intended to generate compassion towardsVincent pragmatic replies indicate that not everyone willdemonstrate compassionate responses in every instanceThe difference between CG and CR Vincents is that be-

ing compassionate towards another being in a conversa-tional narrative context is unlike doing guided exerciseson self-compassion about oneself The frequency of con-structing compassionate replies is a way to practice self-compassion users of CR Vincent spent more time practicingself-compassion than those with CG Vincent Therefore CRVincent was more effective than CG Vincent since CR Vin-cent provided more opportunities to be compassionate Thecaveat is that the link between frequency of practice andincrease in self-compassion may not be direct Althoughmindfulness and self-kindness were most often observedonly common humanity improved significantly accordingto our exploratory quantitative analysis Finding commonhumanity in and through a chatbot is also a strong theme inour thematic analysis

Thematic analysis We categorized our data according tothree pillars of self-compassion [32] as displayed in Table 1While all three sub-components were present in both care-receiving and caregiving conditions more instances occurredwith CR Vincent The numbers below a theme (Table 1) arecounts of how many times it occurred in each condition Allquotes below were to CR VincentAs quotes in Table 1 suggest many participants offered

helpful advice to CR Vincent Vincent showed appreciation

Theme QuoteMindfulnessCaregiving 3Care-receiving 25

There are worse things thatcould happen What has hap-pened has happened

Self-kindnessCaregiving 7Care-receiving 21

Go do something fun todaylike watching a movie Staypositive and keep trying untilyou succeed

Common humanityCaregiving 0Care-receiving 11

Everyone makes mistakesJust remember that it can hap-pen to anyone and that itrsquos notyour fault

Table 1 Self-compassion sub-components

with follow-up statements like ldquoyou always make me feel bet-terrdquo Negative counterparts to three pillars of self-compassionwere not strongly present ie self-judgment was detectedfour times for CR Vincent and once for CG Vincent isolationwas noted once for CR Vincent but none for CG Vincentand over-identification was neither present for CG nor CRVincent

For both conditions people were mostly friendly to Vin-cent and there were no swear words or abusive languagedisplayed The most hostile comment was ldquoyoursquove beenpretty dumbrdquo to CR Vincent and we encountered suchldquoput-downsrdquo only twice There were additional topics thatemerged through open thematic analysis They are sum-marized in Table 2 and these themes could also pertain toself-compassion themes (messages to CGVincent are markedwith ldquoCGrdquo and otherwise they were to CR Vincent)

Theme QuotePragmatismCaregiving 0Care-receiving 41

Maybe next time make a bet-ter planning and make sureyoursquove got enough time )

Perspective-takingCaregiving 0Care-receiving 10

I would find a window toclimb in But maybe in yourcase better try to hack into thefolder [] be proud of thebot that you are

Engagement vsDistantiationCaregiving 27 vs 6Care-receiving 5 vs 11

A girl told me she loves meAnd I love her too (CG) vsSorry itrsquos confidential (CG)

Positive vs NegativeCaregiving 74 vs 9Care-receiving 5 vs 2

I was laying in a field full offlowers trying out my newukelele (CG) vs I hate pink

Table 2 Free-input themes

People interacted with CG and CR Vincents differently(Table 2) Giving pragmatic advice to Vincent and taking itsperspective as a chatbot were themes only found in the CRcondition Rather than tending to Vincent by giving emo-tional support participants gave practical advice on what todo better Examples of perspective-taking are recommendingVincent to ldquohack into the folderrdquo or to use ldquobrute forcerdquo tech-niques to gain access participants thought like a chatbot tohelp a chatbotSome participants revealed more personal information

to CG Vincent (theme engagement) and took interest inVincent by asking questions back eg ldquowhat (did you) dofor money before nowrdquo or writing lengthy responses Someshared information was very intimate in nature eg ldquoIrsquomgoing to kiss the girl next to me in 5 secondsrdquo Since CGVincent asked participants to write about themselves thisskewed engagement (the amount of textual response) to-wards CG Vincent Participants distanced themselves fromVincent only a few times by stating that certain informationwas confidential or not showing interest in getting to knowVincent eg ldquosorry I do not know what interests yourdquo Thelast theme on positive vs negative attitude was primarilypresent in the CG condition this theme was mostly aboutattitudes participants had about themselves and their livesnot about Vincent Most participants shared positive lifeevents eg getting an internship cooking something deli-cious Though negative attitudes were minimal they rangedfrom more mundane states eg feeling ldquoawkwardrdquo to moredramatic states eg ldquonot die within 2 weeksrdquoTo summarize Tables 1 and 2 self-compassion sub- com-

ponents were more present with CR Vincent suggestingthat giving compassion to Vincent (or another being) thantowards oneself may be more natural in conversational con-texts And mindfulness most frequently occurred (Table 1)As for emergent themes in Table 2 participants gave prag-matic advice to CR Vincent and often practiced perspective-taking Yet CG Vincent allowed for more self-expression ifparticipants were open to communicate as shown by greaterinstances of engagement and positive remarks about every-day situations In a few instances we detected deeply per-sonal messages on ups and downs of relationships and self-deprecating thoughts Mostly participants shared positivedaily news with CG Vincent and helpful or uplifting remarkswith CR Vincent

Interpretive analysis We now offer a broader interpreta-tion of our data by incorporating participantsrsquo open-endedresponses to an item on the final survey The main theme isbonding between participants and Vincent though not allbonded with Vincent in the same way To explain this weprovide three subthemes that underlie the bonding processwith Vincent Our primary focus was on CR Vincent

Relatability leads to believability Participantsrsquo ability toextend the sense of common humanity to a chatbot touchesupon anthropomorphism CR Vincent was comforted as ifit were a human eg ldquoitrsquos human to make mistakesrdquo (CR)while its problems were addressed to its ldquochatbot worldrdquo egldquocommunicate whatrsquos going on to your fellow chatbotsrdquo (CR)For one participant even Vincentrsquos limitation of having astrict script was anthropomorphized ie ldquoVincent is like thefriend who always speaks about himself and what he haslearned or done and sometimes out of courtesy (not out ofcuriosity) asks how you are doing - but doesnrsquot listen to youranswer or what you actually have to say he just goes onwith his own thingrdquo (CG) Such attributed anthropomorphictraits depended on participantsrsquo willingness take Vincentrsquosperspective as a chatbotCR Vincentrsquos blunders were based on common human

mishaps like being late for a meeting and dealing with un-paid bills (scenarios from [13]) Yet none of our participantsquestioned whether or not a chatbot had meetings to attendor bills to pay Vincentrsquos narrative was on how a chatbotcould be late (new updates took longer than expected) orhow it could have bills (Vincent needs to pay the hostingserver) and our participants went along with imagined sce-narios Vincent faced Rather than questioning the parametersof our scenarios on realism participants thought of how tosolve Vincentrsquos problems within the parameters of a chat-botrsquos world When relevant CR Vincent played up the ironyof having human struggles as a chatbot eg ldquoall I am is lit-erally a piece of code and I failed a programming courserdquoVincent became believable because its struggles were relat-able Granting Vincent human-likness was less literal in howpeople bonded with Vincent Vincent did not try to appearhuman but it socialized with participants about its strugglesthat humans also had People related to Vincentrsquos strugglesand believed that such struggles could arise for chatbotsShared history can lead to attachment Conversations be-

tween people as well as in human-computer interactionbecome shared history over time For instance ldquo[] commu-nicating with Vincent everyday for two weeks builds somekind of habit It makes me notice its presence and absence(which might be good) I think it has a potential to be a goodcompanion and improve the mood especially if someone isfeeling lonelyrdquo (CG) Thus frequent communication with achatbot in a given duration can form expectations ldquoI reallymissed Vincent when he started our conversation laterdquo (CR)The level of attachment for some participants was higherthan others eg after the experiment we saw reactions suchas ldquocan I keep himrdquo (CG)When Vincent prepared participants for its daily good-

byes eg ldquoI have some chatbot things to do Defragmentmy server stack Buy aluminum foil to make fashionablehats withrdquo what was intended as humor can be interpreted

differently ie server defragmentation could be life-or-deathfor a chatbot Some people can be confused worried or evenangered when a chatbot they care about does not respondThus one reaction was ldquothe asshole decided to delete itsstack and when I said itrsquod die it just didnrsquot reply You canrsquotgo making people worried about a freaking chatbotrdquo (CR)People may miss a chatbot that suddenly leaves them orsincerely worry about its well-being This is the positiveand negative aspect of a relatable a chatbot some partic-ipants found common-humanity in Vincent and of thoseparticipants a few possibly related more through empathicdistress rather than through compassion If two weeks canbring about strong signs of attachment longer periods ofinteraction may heighten the level of attachment to differentdegrees and in different ways per person

Emotional reciprocity with chatbots As mentioned beforemost participants were able to respond to CR Vincentrsquos emo-tional displays on a practical level eg recommending howto fix a problem or advising Vincent on how to adjust itsemotions eg telling Vincent to stay positive To add somepeople may not appreciate chatbots demonstrating feelingsOthers may reciprocate or feel comforted by a chatbotrsquos ex-pressed emotions even if a chatbot is perceived as incapableof having emotions The more nuanced point is that Vin-centrsquos display of emotions was noted to bring conflictingfeelings For instance ldquowhen Vincent would show emotions(for example lsquolove talking to yoursquo lsquomiss yoursquo) that would feelweird because I know I am talking to a chatbot and it proba-bly is not that developed that it does have feelings But theusage of such words does feels nice compared to when a hu-man would say them So I had conflicted feelings about thesekind of expressionsrdquo (CG) The participant felt conflictedabout how to process Vincentrsquos emotional outreach

Importantly the participant suggested heshemay bemorecomfortable with a chatbot saying ldquomiss yourdquo than a humanTo conjecture the participant could mean that there was nosocial pressure due to a chatbot not expecting or needinghimher to say ldquoI miss you toordquo People often feel obligatedto respond to sincere emotions of others with similarly va-lenced emotional displays even if they do not feel the samesincere emotions towards them Such pressure may not holdfor technological entities Perhaps to miss someone impliesa certain history in a relationship so to hear that from aperson one met less than two weeks ago may feel awkwardor insincere whereas a chatbot would not be expected toknow or abide by certain social conventions If two peopleknew beforehand they will get to know each other for a maxi-mum duration of two weeks (as our participants knew beforemeeting Vincent) and never be in touch again their emo-tional performance may adjust accordingly The timescalefor intensifying socially acceptable emotional expressions

in human-chatbot interactions and human-human interac-tions may differ The ldquolifespanrdquo of a chatbot is not equatableto a personrsquos lifespan And the distinction between super-ficial vs genuine emotional displays from and to a chatbotis not entirely equatable to emotions people share and re-ciprocate between each other Currently we do not haveestablished norms on how emotions between humans andbots areshould be managed We suggest there may be dis-tinct differences compared to emotions in human-humanrelationships

Discussion and design implicationsCR Vincent adds depth to the CASA paradigm [16 31]mdash notonly do people treat a chatbot as an anthropomorphizedsocial agent but they themselves are affected by a chatbotto the extent that their self-compassion can increase whenthey are compassionate towards a chatbot Brave et alrsquosinsight on embodied conversational agents is that ldquojust aspeople respond to being cared about by other people usersrespond positively to agents that carerdquo [4 p 174] We addthat just as giving care to another human can increase onersquosself-compassion [5] caring for a chatbot can enhance onersquosown self-compassion If the dominant question has beenldquowhat can technology do for usrdquo Vincent demonstrates thatby exploring ldquowhat can we do for technologyrdquo we inad-vertently benefit from technology potentially more so thanwhen we only shape technology to serve us This claim isspecified to bots in the mental health domain and our goalwas to increase self-compassion as a target for well-being[52] rather than to reduce clinically defined symptoms ofpsychological ill-being We present our design implicationsbelow on building chatbots for psychological health carewhich primarily stem from our interpretive analysis Ourimplications are inter-related starting points that should becontextualized for each research and deployment processGive users more closed-inputs or free-input optionsMany

participants felt limited in responses they could give to Vin-cent They wanted to write to Vincent without having anypreset answers or needed more options A recommendationis to use natural language processing for a chatbot whichwill rely less on a pre-planned narrative arc and build moreon what users say This will require a longer developmentperiod The simpler option is to provide users with morefixed responses (three to four) and more opportunities foropen input

Develop a chatbotrsquos story with users Conversational agentscan be powerful storytellers [25 36] even without complexAI To deliver co-storytelling as shared history with interac-tants we suggest designers to create flexible narrative pa-rameters that people can creatively use to relate to a chatbotVincent was able to tell its story but it was less interactive inthat people could follow along with limited reaction options

due to the nature of our experiment There can be additionalcomplexities that designers can add For instance the nar-rative can take a different route depending on which closedinput options users click on We have utilized a limited num-ber of messages called ldquopathsrdquo (supplementary material) thatVincent could give depending on closed input responsesYet this practice did not change Vincentrsquos narrative Givinga chatbot ldquomemoryrdquo be it knowing basic information likenames or a more involved retention of what users say canenhance conversational storytellingTread carefully with emotional expressionsWe suggest a

broader view on what emotions are by considering inter-related emotions that develop over time For example for abot to miss someone assumes a botrsquos happinessenjoymentexperienced during a prior interaction with a user a botrsquosability to feel longing should follow its prior display of joyshared with the user This requires critically formulating in-tentions behind communicative moves [44] of any affectivebot There are several paths for developing emotional dis-plays To list a few (1) offer one type of consistent emotionalexpressions as Vincent did (2) design emotional expressionsthat may be appropriate for different target groups in tan-dem with the implication below and (3) give users controlover how their chatbots ldquofeelrdquo towards them The caveat forthe third recommendation is that the user control over achatbotrsquos emotions may not aid a chatbotrsquos narrative andit also may not be helpful for all users the associated riskis that user-controlled emotions can render a chatbot lessrelatable More specifically the illusion that a chatbot canauthentically care for or be cared by another being requiressome level of perceived independence in how it ldquofeelsrdquo Werecommend designers to engage with the growing field ofaffective computing [41 45] and its discussion on ethics [8]If a chatbotrsquos goal is bettering usersrsquo psychological statesdesigners must ask if an affective bot delivers the intendedtreatment and what ethical boundaries there are in its dis-plays and elicitation of emotions Designers and users couldcontrol a chatbotrsquos emotions but what emotions a chatbotcan elicit in users is not always a priori foreseeable

Tailor chatbots to different target groups Evenwith one con-struct self-compassion we see a variety of ways a chatbotcan be configured To start people with low self-compassionmay benefit the most from Vincent as our exploratory anal-ysis shows This can mean more compassion focused sce-narios rather than neutral scenarios Women are noted toscore lower on self-compassion [51] yet older women experi-ence greater compassion than older men [29] Chatbots thatconsider gender age andor occupation can be potentiallyhelpful for increasing self-compassion To list a few examplesfor reincarnating Vincent a chatbot could be gendered asfemale or non-binary present a proactive version of compas-sion specified for women (see eg Neff [35] on speaking up

and protecting oneself from harm) talk about exam stresswith students or refer to stressful meetings or workplacebullying with employed individuals Rather than assumingthat one-size-fits-all or extreme personalization will workwe suggest designers to first approach targeted groups toclearly understand their needs For instance whether a self-compassion chatbot for all women is as effective or moreeffective than a more targeted chatbot eg at various levelsof intersectionality like race culture age etc should beconsidered given the time and resources that may be avail-able We recommend that based on research uncoveringpossible ways to design a chatbot that suits different needsand wants should be prioritized

Future works and limitationsOur study opened up new questions to be explored An av-enue to investigate is how a chatbotrsquos use of emotional lan-guage influences its interactants We posit that a sufferingchatbot induces less empathic distress than a suffering hu-man and whether or not this is the case needs to be furtherinvestigated especially for chatbots intended to be therapeu-tic helpers An awareness of onersquos own and othersrsquo sufferingwithout overwhelming empathic distress is suggested to bepossible through compassion [18 47] Hence disambiguat-ing compassion from empathic distress is critical in deploy-ing self-compassion chatbots as instantiations of positivecomputing [6] Different configurations of Vincent basedon peoplersquos gender age or occupation could improve theirself-compassion scores more effectively and if and in whatways this holds true warrants further research

There are limitations to consider Our effect size was lowerthan findings from the Woebot study (d = 044) which had34 participants who ldquoself-identified as experiencing symp-toms of depression and anxietyrdquo [15 p 2] and they measuredsymptoms of depression with the PHQ-9 questionnaire notself-compassion Falconer et alrsquos results on self-compassionscores after embodied VR experience also had a higher ef-fect size with the partial eta-squared of 036 (d = 1487) [14]which was based on 15 participants with depression Weworked with a general non-clinical sample and CG Vincentshowed an effect size of d = 007 (N = 34) and CR Vincentrsquoseffect size was d = 02 (N =33) Follow-up studies on self-compassion chatbots can utilize a larger sample and a moregrounded effect size One explanation for the difference ineffect size is that we did not recruit people who were clini-cally or self-proclaimed to be depressed based on the viewthat preventative mental health care can build resilience forpeople in general While Vincent andWoebot [15] share com-monalities the main measurements and targeted populationdiffered And while self-compassion was the measurementfor us and Falconer et al [14] the technology used sample

size and targeted population differed The gain andor main-tenance of healthy self-compassion as pre-emptive care maynot result in a similarly high effect size but can be psycholog-ically beneficial nonetheless More research is necessary tounderstand long-term consequences of a priori preventativecare vs a posteriori treatment of mental health woes

More broadly peoplersquos engagement with Vincent may re-flect both socially desirable reactions such as politeness to-wards machines as social actors [16 31] as well as emotionalempathy ie the ability to ldquofeel forrdquo VincentWe have not yetconcretely looked into other potential contributing factorsin bringing about self-compassion through human-chatbotinteraction Also what is difficult to gauge is the magnitudeof a chatbotrsquos perceived social and emotional complexitybased solely on messaging or text-based conversations Vin-cent lacked embodied communication which means it didnot use non-verbal modalities such as gaze voice or ges-tures that are critical in various social interactions Vincentwas a uni-modal technological entity that can be extendedthrough other complex emotional displays Thus we havenot established how people would engage with other formsof technology like robots with varying degrees and types ofembodiment alongside different combinations of modalitiesUtilizing technology appropriately for mental health carerequires many comparative renditions

5 CONCLUSIONCompassion is a key moral emotion [19] or motivation [6]that deserves to be further explored through positive comput-ing or technology for well-being Self-compassion can helppeoplersquos overall well-being [52] through kindness towardsoneself connectedness to greater humanity andmindfulnessWhile a chatbot is not a panacea for curing psychologicaldifficulties and is not meant to replace professional helpwe demonstrated that caring for a chatbot can help peoplegain greater self-compassion than being cared for by a chat-bot Our quantitative and qualitative analyses suggest thathuman-chatbot interaction is a promising arena for positivecomputing

REFERENCES[1] Anxiety and Depression Association of America (ADAA) 2016 ADAA

Reviewed Mental Health Apps httpsadaaorgmental-health-appsAccessed 2018-06-25

[2] Arthur Aron Elaine N Aron and Danny Smollan 1992 Inclusion ofOther in the Self Scale and the structure of interpersonal closenessJournal of Personality and Social Psychology 63 4 (1992) 596 httpsdoiorg1010370022-3514634596

[3] Virginia Braun and Victoria Clarke 2006 Using thematic analysisin psychology Qualitative Research in Psychology 3 2 (2006) 77ndash101httpsdoiorg1011911478088706qp063oa

[4] Scott Brave Clifford Nass and Kevin Hutchinson 2005 Computersthat care investigating the effects of orientation of emotion exhibited

onersquos suffering with understanding connectedness over isola-tion is to view onersquos life as intertwined with other lives ratherthan to see onersquos experiences as unrelated or irrelevant togreater humanity mindfulness over over-identification is tobe aware of onersquos negative emotions in a balanced mannerthan to excessively identify with them [52]Onersquos gender may influence self-compassion A meta-

analysis concluded that women score lower than men onself-compassion and the gender difference was especiallypronounced when sampled studies had more ethnic minori-ties [51] Women reportedly have greater empathy than men[23] and they are more likely to be more self-critical thanmen [51] To add women who provide empathy as socialsupport can feel drained or distressed [22] Yet a study witholder adults demonstrated that older women have greatercompassion than older men [29] While peoplersquos gender ageand possibly ethnic minority status may impact their self-compassion practicing self-kindness connectedness andmindfulness can help individuals be more compassionatetowards themselves and others

In clinical settings people who experiencemental ill-beingcan benefit from self-compassion [17] Self-compassion isalso strongly connected to well-being for the general popu-lation [52] Well-being refers to mostly feeling more positiveaffect than negative affect and being satisfied with onersquos lifefactors like income gender age or culture influence onersquoswell-being only minutely [30] Thus having a good balancebetween onersquos psychological social and physical capabili-ties to deal with lifersquos difficulties is important for well-beingrather than having a static life without suffering [12] (noris this realistic) Through an awareness of onersquos and othersrsquosuffering without being overwhelmed by empathic distresscompassion is developed [18 47] Caring for or being com-passionate towards others has been shown to increase onersquosown self-compassion [5] Yet could the same effect be foundwhen people care for technological entities Technology canpotentially be a means to achieve self-compassion and byextension well-beingWe return to the question ldquowhen bots have psychologi-

cal issues can humans care for them and if so howrdquo toemphasize that (1) anthropomorphic realism of artificial en-tities is not required for humans to develop a caretakingstance towards them and (2) machinesrsquo mimicry of peoplersquospsychological ill-being can help ascertain why certain psy-chological traits are labeled as issues When we observe howpeople take care of unwell chatbots we may uncover howthey themselves want to be treated Machines therefore donot need to pass the Turing test for the purpose of positivecomputing or technology for well-being [6]

In order to uncover peoplersquos psychological responses tochatbots particularly in relation to modulating peoplersquos self-compassion we asked participants to interact with a chat-bot Vincent designed to engage in care-giving versus care-receiving conversations Exploring simulated psychologicalstates via technological entities like Vincent is a way to en-vision positive computing In addition our approach focuseson pre-emptive mental health care We now turn to howwe designed our study how we built the two Vincents andpresent our results

3 METHODWe utilized quantitative and qualitative methods to best un-derstand our data We compared self-compassion scores be-fore and after two weeks of interaction and examined if theCR and CG conditions showed any difference Our mixedlongitudinal design was supplemented by thematic [3] andinterpretive analyses [48]Our qualitative analysis was performed on participantsrsquo

open-ended responses to Vincentrsquos questions eg ldquocan youremember something that went well for you recentlyrdquo (CGVincent) ldquocan you think of something that will cheer me upa bitrdquo (CR Vincent) and on post-experiment open endedresponses about the experience in general We coded deduc-tively on the sub-scales of the self-compassion scale and weallowed for inductive themes to emerge [3 48] Four codersanalyzed the data and a fifth annotator broadly checkedfor coherence or incoherence resulting in a structured it-erative process Our quantitative and qualitative measure-ments therefore corresponded with each other to triangulatevarying insights of the same phenomenonmdash self-compassionthrough human-chatbot interaction

Participants and groupsWe conducted a power analysis to estimate our sample sizeOur effect size is based on an aforementioned study [14] thatmeasured self-compassion of a clinical population (N = 15)in an embodied VR experiment with the partial eta-squaredof 036 at p = 002 [14] which gave us a sample size of 68with a power of 90 and an error probability rate of 005 Weplanned for t-tests and thus the transformed effect size viaeta-squared to Cohenrsquos d was 1487 which was reduced toa more realistic 08 We had 67 participants (F = 29 M = 36undisclosed = 2) with the mean age at 251 years (SD = 57range = 19 - 48)We recruited people through the participantsdatabase of the Eindhoven University of Technology Allparticipants completed the self-compassion questionnairebefore they interacted with Vincent Then they were dividedto two conditions so that the average self-compassion scoreswere evenly distributed at the start From the lowest scoringto the highest scoring participants we divided all into eitherthe CR or the CG condition in an alternating manner This

practice resulted in a relatively even gender distribution forboth conditions (CR M = 18 F = 14 undisclosed = 1 CG M= 18 F = 15 undisclosed = 1)

Chatbot implementation

Figure 1 Care-receiving Vincent

Vincent2 was built with Googlersquos Dialog flow that wasintegrated to Facebook Messenger3 We purposefully did notvisually design Vincent (the profile icon showed a ldquoVrdquo) todrive Vincentrsquos personality by what was said rather thanhow Vincent looked Participantsrsquo responses did not changeVincentrsquos reactions The usage of limited pre-set responses(Figure 1) was to continue Vincentrsquos narrative whilst allow-ing participants a choice between relevant responses (Woe-bot also interacted with its users in a similar fashion [15])This allowed for greater comparability between participantsin each condition

We had eight scenarios each for CG or CR Vincent and 6neutral scenarios (total 22 scenarios) The neutral scenarioswere the same for both Vincents and aimed for adherenceto daily touch-points in an entertaining way For exampleone neutral scenario is about world records ldquo[] I knew you

2Vincentrsquos Facebook page - httpswwwfacebookcomvincentthebot3Dialog Flow Facebook Integration - httpsdialogflowcomdocsintegrationsfacebook

were an interesting bunch of carbon-based beings but appar-ently you have this thing called lsquoworld recordsrsquo [] [] whatwould you like to be in the world records book forrdquo BothVincents used images and emojis because visual icons arewidespread in digital messaging to express emotions [43]Weused appropriate punctuation marks and positively and neg-atively valenced syntax in accordance to previous researchon text-based emotion expression and detection [20] Mostinputs were close-ended (Figure 1) but we allowed open-ended free inputs at least once per interaction We aimedfor a meaningful comparison in two ways In each condi-tion participantsrsquo interactions were designed to be the samewith limited set of possible reactions to Vincent Betweenthe two Vincents we wanted to clearly distinguish betweenthe recipient and the giver of care Below are excerpts fromCG and CR Vincents (all scenarios are in the supplementarymaterial)

CG Vincent [] see if you can think of a kinder more caringway to motivate yourself to make a change if needed Whatis the most supportive message you can think of thatrsquos in linewith your underlying wish to be healthy and happy Try towrite it belowUser [free input]

CR Vincent What do you think am I the dumbest bot yoursquoveever seen or whatU [free input]V Am I being too hard on myselfU [free input]

CG Vincent was modeled after Woebot [15] supplementedby self-compassion exercises [34] CR Vincent was based onthe Self-Compassion and Self-Criticism scale [13 14] Thescale has scenarios like job rejections unpaid bill remindersand being late to a meeting which were converted to fita chatbot (Figure 1) The eight items of the scale were in-terweaved for conversational storytelling By doing so wejuxtaposed issues that students (the majority of our sample)can face eg distress about failing an exam that Vincentunderwent CR Vincent narrated its story over time by ad-mitting its mistakes feeling inadequate compared to otherchatbots confessing a former lie and asking for confiden-tiality in sharing a non-trivial worryI got a reminder from the server that hosts me Itrsquos like my

house so to say [] I forgot to pay the server fee on time[][] It wouldrsquove taken me only 0004 seconds to make the trans-action you know since Irsquom a robot and all [] this never seemsto happen to the other chatbots on my serverVincent brought up this scenario again later with new

informationRemember our talk a couple of days ago About me forgetting

to pay my server fee in time [] I kind of lied to you [] Ididnrsquot tell you this before because I was a little embarrassedabout it Can you promise me that this stays between us []Irsquove been applying for different jobs and just today I receivedmy third rejection email already The reason that I couldnrsquotpay my bill was because Irsquom running out of money And if Idonrsquot find a job soon Irsquoll get kicked off my server

In sum CG Vincent guided participants through activitieswhile CR Vincent opened up about its everyday mistakesor worries to get support from participants CG Vincentsought to trigger self-compassion in users themselves and CRVincent gave participants opportunities to be compassionatetowards a chatbot

MeasurementsWe collected basic demographic information ie genderage and previous experience with a chatbot at the startOur main measurement was the self-compassion scale on afive point scale [32] with six sub-components (self-kindnessvs self-judgment common humanity vs isolation mind-fulness vs over-identification) which was deployed twicebefore and after the experiment We added a measurementfor opinions about the agent on a seven point scale [4] todetect irregularities between how CG and CR Vincents maybe perceived Our final scale was the Inclusion of Other inthe Self (IOS) Scale a single item on a seven point scale [2]to check howmuch participants identified with Vincent post-hoc We also kept track of two additional aspects to gaugeengagement quality One is the error rate ie the number oftimes Dialogflow crashed during the interaction sometimesrequiring an experimenter to momentarily ldquowizardrdquo [9] torestart the interaction4 The other is the total word countper participant on open ended answers

ProcedureWe built Vincent and wrote our initial scenarios to be testedOur pilot study of three days was with voluntary participants(N = 12) personally recruited by experimenters We checkedif scenario categories (caregiving care-receiving and neu-tral) were clear by asking participant to guess Vincentrsquosintentions and goals per scenario Based on this we onlyadapted neutral scenarios Then we recruited participantsfor the actual experiment Our email invitation was sent outto the TU Eindhoven participant database and interestedparticipants joined the experiment if they used Facebookand could freely express themselves in English The emailcontained a link to Vincentrsquos Facebook page that partici-pants had to go to for a guided tutorial on the experimentpayment information and the informed consent form This

4Restarts happened 37 out of 938 interactions (14 days 67 participants) or394 of the time

form noted that experimenters will look at participantsrsquo dataand that third party technology providers Vincent relied onie Facebook and Google have access to the informationWe added that participantsrsquo personally identifiable informa-tion will not be shared for publication purposes and thattheir voluntary participation means that they can drop outof the study at any point We also stated that Vincent is nota therapist and that they should seek professional help if anypsychological issues are experienced though we targeted anon-clinical population

After the guided tutorial participants filled in the first setof questions on demographic information and self-compassionThen they were assigned to either CG or CR Vincent Theyall began the experiment on the same day For two weeksVincent greeted and sent a password daily and participantshad to repeat the password to Vincent to start the daily inter-action eg Hey me again ) Tell me tHup to start our littletalk After two weeks participants filled in the final surveyon self-compassion the IOS scale opinion on the chatbotdetails for compensation as well as additional comments orfeedback they were willing to share Participants were thenpaid through bank transfer

4 RESULTSWe first present our quantitative analysis and then move onto how people talked with Vincent the qualitative angle Allrelevant tables and graphs as well as CG and CR Vincentsrsquoscenarios are in the supplementary material

Quantitative analysisBefore we forged ahead with hypotheses testing we lookedinto engagement levels of all participants to detect outliersWe had three outliers participants who had less than 20minutes of total interaction time with Vincent Only readingwhat Vincent sent not including giving a response shouldtake one to two minutes per day We expected a minimum of20 to 28 minutes of interaction for two weeks (for all partici-pants the average total time was 36 minutes SD = 10) Ouroutliers spent in total 15 18 and 19 minutes each the threelowest total interaction time Correspondingly their totalword count to open responses reflected low engagement at27 22 and 27 words for the total duration of the experimentthe three lowest total word count out of all participants

We conducted two one-tailed dependent t-tests [7] to an-swer our hypotheses (we set p at 005 with the confidenceinterval of 95) CG Vincent did not result in significantchange (t(31) = -0572 p = 0286 Cohenrsquos d = 007) whencomparing before (M = 3135 SD = 0630) and after (M =3180 SD = 0628) the two weeks but the direction detectedis positive (normality assumed for prior (W = 0987 p = 0978)and post (W = 0989 p = 0982) scores) CR Vincent did showa significant difference (t(31)= -197 p = 0029 Cohenrsquos d

= 02) between prior (M = 3137 SD = 0613) and post (M= 3257 SD = 0558) scores for self-compassion (normalityassumed for prior (W = 0979 p = 0764) and post (W = 0946p = 0114) scores)We conducted exploratory analyses to better understand

our data Through a repeated measures ANOVA we checkedfor the effect of time prior and post self-compassion scores(F(1 62) = 2768 p = 0101 η2 = 0043) Then we checked forthe effect of condition ie CG or CR (F(1 62) = 075 p = 0785η2 = 0001) and the interaction between time and condition(F(1 62) = 0580 p = 0449 η2 = 0009) None were significantWe additionally checked for time condition timeconditioneffects on the three components of self-compassion self-kindness common humanity and mindfulness Only theeffect of time on common humanity was significant F(1 62)= 6059 p = 0017) The table for all statistical tests is in thesupplementary materialWe further dissected our data by gender because previ-

ous research showed that women may score lower on self-compassion [51] Indeed female participants had lower self-compassion scores (M = 305 SD = 013) than men (M = 326SD = 009) at the start but not significantly so (t (4935) =113 p = 026 d = 029) according an independent unequalvariance t-test (normality assumed) We then compared postand prior scores for men and women Menrsquos self compassionscores increased only by 002 as a difference in means andshowed no significant increase (t (33) = -025 p = 040 d =004) However womenrsquos scores showed a significant differ-ence (t (27)= -206 p = 002) between 305 (SD = 071) as thestarting score and 319 (SD = 065) as a posterior score forself-compassion When we scrutinized the gender differencebetween CG and CR Vincents we noticed a more dramaticdifference Women with CR Vincent showed a highly signifi-cant change (t(13) = -289 p = 0006 d = 077) compared towomen with CG Vincent (t(13) = -033 p = 037 d = 009)and both were normally distributed

We wanted to check if mainly gender was at stake or if itwas simply a difference between low vs high scorers on priorself-compassion levels We thus divided all participants intotwo groups based on the average self-compassion score at thestart 314 Those who scored above this were high scorers(M = 371 SD = 037 N = 28) those who scored below werelow scorers (M = 269 SD = 032 N = 36) We included oneparticipant with the average score in the high-scoring groupand this had no impact on significance reached Low scorersgreatly increased their self-compassion scores in terms ofsignificance (t(35) = -341 p = 00008 d = 057) but highscorers did not show improvements (t(27) = 110 p = 086 d= 018) Yet normality was not assumed for both low-scorers(W = 093 p = 003) and high-scorers (W = 091 p = 002)since we divided a normally distributed group into two Thuswe performed a two-sample Wilcoxon rank-sum test to see

if there was a significant difference between low scorers andhigh scorers which was the case at z = 286 and p = 0004 CRVincent improved low scorersrsquo self-compassion significantly(t(17) = -320 p = 0003 d = 075) compared to a marginalsignificance for CG Vincent (t(17) = -175 p = 005 d = 041)There were normal distributions observed for prior and postself-compassion levels of low and high scorersA potential explanation for why low scorers improved

more than high-scorers is regression to the mean Howeverin published literature the average self-compassion score isbetween 25 and 35 [33] and our low scorers have a prior av-erage self-compassion of 269 If regression to the mean is anexplanation we would also expect high-scorers to end witha lower mean yet this is not the case Our high-scorers hadan average prior score of 371 (above average [33]) and theirscores did not decrease after the experiment This may be aceiling effect The low-scorersrsquo improvement is still there itis a highly significant effect even with the Bonferroni correc-tion for all tests (p = 00008) with the post-hoc power of 097The data supports that Vincent enhanced self-compassionfor low-scorersWe had two additional scales one to check if people per-

ceived CG and CR Vincent in a relatively similar way [4]and the other to see how much participants identified withVincent [2] The survey on participantsrsquo opinion of the agentincluded caring likability trustworthiness intelligence dom-inance and submissiveness as items [4] about Vincent BothCG and CR Vincents were perceived to be fairly analogoussave for dominance (α = 048 CG Vincent M = 2677 SD =0794 CR Vincent M = 2656 SD = 0700) and submissiveness(α = 024 CG Vincent M = 2448 SD = 0559 CR Vincent M= 2396 SD = 0636) though none showed a significant dif-ference between two conditions The the Inclusion of Otherin the Self (IOS) scale [2] displayed that participants moreclosely related to CR Vincent (M = 348 SD = 148) thanCG Vincent (M = 306 SD = 139) but not significantly so (t= -115 p = 025 d = 029) Both conditions were normallydistributed (CG W = 096 p = 034 CR W = 097 p = 064)

To summarize our hypothesis that CG Vincent increasesself-compassion was not supported (p = 0286) but the hy-pothesis that CR Vincent increases self-compassion was sup-ported (p = 0029) Our exploratory analyses captured threeunderlying influences on this finding First our ANOVA testsrevealed that the only significant aspect was time as an inde-pendent variable affecting common humanity one elementof self-compassion (p = 0017) Second gender may be a con-tributing factor with women demonstrating a significantincrease in self-compassion (p = 002) for both conditionscombined but not men (p = 040) To add only CR Vincentdemonstrated a highly significant change for women (p =0006) unlike women who interacted with CG Vincent (p =037) Third regardless of gender those who started out with

a low self-compassion score exhibited the most significantchange (p = 00008) for both conditions together Low-scorersmore significantly improvedwith CR Vincent (p = 0003) thanwith CG Vincent (p = 005)

Put together CR Vincent more effectively increased self-compassion than CG Vincent most likely through a signif-icant change in participantsrsquo sense of common humanitymore so than self-kindness and mindfulness (as our ANOVAtest showed) Finding common humanity can be inclusive ofchatbots Women specifically those with CR Vincent weresignificantly more affected than men However low-scorersof both genders benefited the most compared to high-scorersespecially those with CR Vincent CG and CR Vincents werenot perceived to be significantly different except for a lowersimilarity regarding the dominance-submissive trait Par-ticipants in the CR condition may have felt that Vincent ismore like them (ISO scale) though this difference was notsignificant

Qualitative analysisFor our qualitative analysis we used the corpus of free re-sponses that participants had typed during their interactionswith Vincent We will first present a descriptive analysisof the interactions that people had with CG and CR Vin-cents followed by our thematic analysis [5] and interpretiveanalysis [48] Participantsrsquo responses to CR Vincent were onaverage 9353 words (SD = 4796) and to CG Vincent 11247words (SD = 6399) for two weeks While our data set is notabundant in terms of word count we believe a qualitativelook at how participants interacted with Vincent is valuable

Descriptive analysis CGVincent guided participants throughself-compassion exercises [34] that did not require them toactively voice aspects of self-compassion they mainly hadto read about it This resulted in fewer instances of self-compassion themes in CG Vincent since they only occurredwhen Vincent asked participants to comfort themselves Toadd participantsrsquo willingness to engage with CG Vincentrsquosprobes differed It asked specific questions or provided ashort task for free input eg ldquowrite down a difficulty youhaverdquo Many answers to this were short ldquowake up earlyrdquoor ldquoI often feel alonerdquo Some participants opened up moreldquoI have a difficulty in expressing myself when I am underdifficult situationsrdquo or ldquoI am studying abroad far from homeand family and friends Different culture language edu-cational standardrdquo CG Vincent asked a follow-up questionldquohow did it make you feelrdquo we again got simple answerslike ldquogoodrdquo or ldquonormalrdquo or longer expressions ldquoI feel morerefreshedrdquo or ldquonot really better or worserdquo In other instancesCG Vincent allowed participants to give themselves simpleself-assurance ldquoI can do itrdquo plan for long-term goals ldquoonceI have graduated I can schedule a good routine to target a

fitter and healthier lifestyle Just hang in there a little longerrdquoor dig deeper ldquopeople around me would be happier if I washappierrdquo Thus CG Vincent provided a daily touch point forself-reflection admittance of everyday suffering ldquopep-talkrdquoor deeper self-insight which may or may not directly relateto self-compassion for all participants

In contrast CR Vincent frequently asked for help and con-sequently received many self-compassion related answersThe narrative was the focus for CR Vincent It was able tobecome more vulnerable over time by admitting its own ev-eryday hardships as a chatbot which led it to seek opinion oradvice For example CR Vincent asked ldquowhat do you thinkam I the dumbest bot yoursquove ever seen or what Am I beingtoo hard on myselfrdquo To this participants responded in dif-ferent ways ldquoI think yoursquore the funniest bot that Irsquove everseen mdash yes you are in some situationsrdquo ldquoNo but you shouldexpect (that) a bot is very smartly programmed and know allmdash Maybe I do not knowrdquo or ldquothe world needs bots like youAnd itrsquos usual to get rejected sometimes just keep on goingand yoursquoll find a job soon enoughrdquo However CR Vincentrsquoscries for help did not always result in necessarily compassion-ate replies Many users stuck to pragmatic answers relatedto the topic of the problem Even though all of CR Vincentrsquosscenarios were intended to generate compassion towardsVincent pragmatic replies indicate that not everyone willdemonstrate compassionate responses in every instanceThe difference between CG and CR Vincents is that be-

ing compassionate towards another being in a conversa-tional narrative context is unlike doing guided exerciseson self-compassion about oneself The frequency of con-structing compassionate replies is a way to practice self-compassion users of CR Vincent spent more time practicingself-compassion than those with CG Vincent Therefore CRVincent was more effective than CG Vincent since CR Vin-cent provided more opportunities to be compassionate Thecaveat is that the link between frequency of practice andincrease in self-compassion may not be direct Althoughmindfulness and self-kindness were most often observedonly common humanity improved significantly accordingto our exploratory quantitative analysis Finding commonhumanity in and through a chatbot is also a strong theme inour thematic analysis

Thematic analysis We categorized our data according tothree pillars of self-compassion [32] as displayed in Table 1While all three sub-components were present in both care-receiving and caregiving conditions more instances occurredwith CR Vincent The numbers below a theme (Table 1) arecounts of how many times it occurred in each condition Allquotes below were to CR VincentAs quotes in Table 1 suggest many participants offered

helpful advice to CR Vincent Vincent showed appreciation

Theme QuoteMindfulnessCaregiving 3Care-receiving 25

There are worse things thatcould happen What has hap-pened has happened

Self-kindnessCaregiving 7Care-receiving 21

Go do something fun todaylike watching a movie Staypositive and keep trying untilyou succeed

Common humanityCaregiving 0Care-receiving 11

Everyone makes mistakesJust remember that it can hap-pen to anyone and that itrsquos notyour fault

Table 1 Self-compassion sub-components

with follow-up statements like ldquoyou always make me feel bet-terrdquo Negative counterparts to three pillars of self-compassionwere not strongly present ie self-judgment was detectedfour times for CR Vincent and once for CG Vincent isolationwas noted once for CR Vincent but none for CG Vincentand over-identification was neither present for CG nor CRVincent

For both conditions people were mostly friendly to Vin-cent and there were no swear words or abusive languagedisplayed The most hostile comment was ldquoyoursquove beenpretty dumbrdquo to CR Vincent and we encountered suchldquoput-downsrdquo only twice There were additional topics thatemerged through open thematic analysis They are sum-marized in Table 2 and these themes could also pertain toself-compassion themes (messages to CGVincent are markedwith ldquoCGrdquo and otherwise they were to CR Vincent)

Theme QuotePragmatismCaregiving 0Care-receiving 41

Maybe next time make a bet-ter planning and make sureyoursquove got enough time )

Perspective-takingCaregiving 0Care-receiving 10

I would find a window toclimb in But maybe in yourcase better try to hack into thefolder [] be proud of thebot that you are

Engagement vsDistantiationCaregiving 27 vs 6Care-receiving 5 vs 11

A girl told me she loves meAnd I love her too (CG) vsSorry itrsquos confidential (CG)

Positive vs NegativeCaregiving 74 vs 9Care-receiving 5 vs 2

I was laying in a field full offlowers trying out my newukelele (CG) vs I hate pink

Table 2 Free-input themes

People interacted with CG and CR Vincents differently(Table 2) Giving pragmatic advice to Vincent and taking itsperspective as a chatbot were themes only found in the CRcondition Rather than tending to Vincent by giving emo-tional support participants gave practical advice on what todo better Examples of perspective-taking are recommendingVincent to ldquohack into the folderrdquo or to use ldquobrute forcerdquo tech-niques to gain access participants thought like a chatbot tohelp a chatbotSome participants revealed more personal information

to CG Vincent (theme engagement) and took interest inVincent by asking questions back eg ldquowhat (did you) dofor money before nowrdquo or writing lengthy responses Someshared information was very intimate in nature eg ldquoIrsquomgoing to kiss the girl next to me in 5 secondsrdquo Since CGVincent asked participants to write about themselves thisskewed engagement (the amount of textual response) to-wards CG Vincent Participants distanced themselves fromVincent only a few times by stating that certain informationwas confidential or not showing interest in getting to knowVincent eg ldquosorry I do not know what interests yourdquo Thelast theme on positive vs negative attitude was primarilypresent in the CG condition this theme was mostly aboutattitudes participants had about themselves and their livesnot about Vincent Most participants shared positive lifeevents eg getting an internship cooking something deli-cious Though negative attitudes were minimal they rangedfrom more mundane states eg feeling ldquoawkwardrdquo to moredramatic states eg ldquonot die within 2 weeksrdquoTo summarize Tables 1 and 2 self-compassion sub- com-

ponents were more present with CR Vincent suggestingthat giving compassion to Vincent (or another being) thantowards oneself may be more natural in conversational con-texts And mindfulness most frequently occurred (Table 1)As for emergent themes in Table 2 participants gave prag-matic advice to CR Vincent and often practiced perspective-taking Yet CG Vincent allowed for more self-expression ifparticipants were open to communicate as shown by greaterinstances of engagement and positive remarks about every-day situations In a few instances we detected deeply per-sonal messages on ups and downs of relationships and self-deprecating thoughts Mostly participants shared positivedaily news with CG Vincent and helpful or uplifting remarkswith CR Vincent

Interpretive analysis We now offer a broader interpreta-tion of our data by incorporating participantsrsquo open-endedresponses to an item on the final survey The main theme isbonding between participants and Vincent though not allbonded with Vincent in the same way To explain this weprovide three subthemes that underlie the bonding processwith Vincent Our primary focus was on CR Vincent

Relatability leads to believability Participantsrsquo ability toextend the sense of common humanity to a chatbot touchesupon anthropomorphism CR Vincent was comforted as ifit were a human eg ldquoitrsquos human to make mistakesrdquo (CR)while its problems were addressed to its ldquochatbot worldrdquo egldquocommunicate whatrsquos going on to your fellow chatbotsrdquo (CR)For one participant even Vincentrsquos limitation of having astrict script was anthropomorphized ie ldquoVincent is like thefriend who always speaks about himself and what he haslearned or done and sometimes out of courtesy (not out ofcuriosity) asks how you are doing - but doesnrsquot listen to youranswer or what you actually have to say he just goes onwith his own thingrdquo (CG) Such attributed anthropomorphictraits depended on participantsrsquo willingness take Vincentrsquosperspective as a chatbotCR Vincentrsquos blunders were based on common human

mishaps like being late for a meeting and dealing with un-paid bills (scenarios from [13]) Yet none of our participantsquestioned whether or not a chatbot had meetings to attendor bills to pay Vincentrsquos narrative was on how a chatbotcould be late (new updates took longer than expected) orhow it could have bills (Vincent needs to pay the hostingserver) and our participants went along with imagined sce-narios Vincent faced Rather than questioning the parametersof our scenarios on realism participants thought of how tosolve Vincentrsquos problems within the parameters of a chat-botrsquos world When relevant CR Vincent played up the ironyof having human struggles as a chatbot eg ldquoall I am is lit-erally a piece of code and I failed a programming courserdquoVincent became believable because its struggles were relat-able Granting Vincent human-likness was less literal in howpeople bonded with Vincent Vincent did not try to appearhuman but it socialized with participants about its strugglesthat humans also had People related to Vincentrsquos strugglesand believed that such struggles could arise for chatbotsShared history can lead to attachment Conversations be-

tween people as well as in human-computer interactionbecome shared history over time For instance ldquo[] commu-nicating with Vincent everyday for two weeks builds somekind of habit It makes me notice its presence and absence(which might be good) I think it has a potential to be a goodcompanion and improve the mood especially if someone isfeeling lonelyrdquo (CG) Thus frequent communication with achatbot in a given duration can form expectations ldquoI reallymissed Vincent when he started our conversation laterdquo (CR)The level of attachment for some participants was higherthan others eg after the experiment we saw reactions suchas ldquocan I keep himrdquo (CG)When Vincent prepared participants for its daily good-

byes eg ldquoI have some chatbot things to do Defragmentmy server stack Buy aluminum foil to make fashionablehats withrdquo what was intended as humor can be interpreted

differently ie server defragmentation could be life-or-deathfor a chatbot Some people can be confused worried or evenangered when a chatbot they care about does not respondThus one reaction was ldquothe asshole decided to delete itsstack and when I said itrsquod die it just didnrsquot reply You canrsquotgo making people worried about a freaking chatbotrdquo (CR)People may miss a chatbot that suddenly leaves them orsincerely worry about its well-being This is the positiveand negative aspect of a relatable a chatbot some partic-ipants found common-humanity in Vincent and of thoseparticipants a few possibly related more through empathicdistress rather than through compassion If two weeks canbring about strong signs of attachment longer periods ofinteraction may heighten the level of attachment to differentdegrees and in different ways per person

Emotional reciprocity with chatbots As mentioned beforemost participants were able to respond to CR Vincentrsquos emo-tional displays on a practical level eg recommending howto fix a problem or advising Vincent on how to adjust itsemotions eg telling Vincent to stay positive To add somepeople may not appreciate chatbots demonstrating feelingsOthers may reciprocate or feel comforted by a chatbotrsquos ex-pressed emotions even if a chatbot is perceived as incapableof having emotions The more nuanced point is that Vin-centrsquos display of emotions was noted to bring conflictingfeelings For instance ldquowhen Vincent would show emotions(for example lsquolove talking to yoursquo lsquomiss yoursquo) that would feelweird because I know I am talking to a chatbot and it proba-bly is not that developed that it does have feelings But theusage of such words does feels nice compared to when a hu-man would say them So I had conflicted feelings about thesekind of expressionsrdquo (CG) The participant felt conflictedabout how to process Vincentrsquos emotional outreach

Importantly the participant suggested heshemay bemorecomfortable with a chatbot saying ldquomiss yourdquo than a humanTo conjecture the participant could mean that there was nosocial pressure due to a chatbot not expecting or needinghimher to say ldquoI miss you toordquo People often feel obligatedto respond to sincere emotions of others with similarly va-lenced emotional displays even if they do not feel the samesincere emotions towards them Such pressure may not holdfor technological entities Perhaps to miss someone impliesa certain history in a relationship so to hear that from aperson one met less than two weeks ago may feel awkwardor insincere whereas a chatbot would not be expected toknow or abide by certain social conventions If two peopleknew beforehand they will get to know each other for a maxi-mum duration of two weeks (as our participants knew beforemeeting Vincent) and never be in touch again their emo-tional performance may adjust accordingly The timescalefor intensifying socially acceptable emotional expressions

in human-chatbot interactions and human-human interac-tions may differ The ldquolifespanrdquo of a chatbot is not equatableto a personrsquos lifespan And the distinction between super-ficial vs genuine emotional displays from and to a chatbotis not entirely equatable to emotions people share and re-ciprocate between each other Currently we do not haveestablished norms on how emotions between humans andbots areshould be managed We suggest there may be dis-tinct differences compared to emotions in human-humanrelationships

Discussion and design implicationsCR Vincent adds depth to the CASA paradigm [16 31]mdash notonly do people treat a chatbot as an anthropomorphizedsocial agent but they themselves are affected by a chatbotto the extent that their self-compassion can increase whenthey are compassionate towards a chatbot Brave et alrsquosinsight on embodied conversational agents is that ldquojust aspeople respond to being cared about by other people usersrespond positively to agents that carerdquo [4 p 174] We addthat just as giving care to another human can increase onersquosself-compassion [5] caring for a chatbot can enhance onersquosown self-compassion If the dominant question has beenldquowhat can technology do for usrdquo Vincent demonstrates thatby exploring ldquowhat can we do for technologyrdquo we inad-vertently benefit from technology potentially more so thanwhen we only shape technology to serve us This claim isspecified to bots in the mental health domain and our goalwas to increase self-compassion as a target for well-being[52] rather than to reduce clinically defined symptoms ofpsychological ill-being We present our design implicationsbelow on building chatbots for psychological health carewhich primarily stem from our interpretive analysis Ourimplications are inter-related starting points that should becontextualized for each research and deployment processGive users more closed-inputs or free-input optionsMany

participants felt limited in responses they could give to Vin-cent They wanted to write to Vincent without having anypreset answers or needed more options A recommendationis to use natural language processing for a chatbot whichwill rely less on a pre-planned narrative arc and build moreon what users say This will require a longer developmentperiod The simpler option is to provide users with morefixed responses (three to four) and more opportunities foropen input

Develop a chatbotrsquos story with users Conversational agentscan be powerful storytellers [25 36] even without complexAI To deliver co-storytelling as shared history with interac-tants we suggest designers to create flexible narrative pa-rameters that people can creatively use to relate to a chatbotVincent was able to tell its story but it was less interactive inthat people could follow along with limited reaction options

due to the nature of our experiment There can be additionalcomplexities that designers can add For instance the nar-rative can take a different route depending on which closedinput options users click on We have utilized a limited num-ber of messages called ldquopathsrdquo (supplementary material) thatVincent could give depending on closed input responsesYet this practice did not change Vincentrsquos narrative Givinga chatbot ldquomemoryrdquo be it knowing basic information likenames or a more involved retention of what users say canenhance conversational storytellingTread carefully with emotional expressionsWe suggest a

broader view on what emotions are by considering inter-related emotions that develop over time For example for abot to miss someone assumes a botrsquos happinessenjoymentexperienced during a prior interaction with a user a botrsquosability to feel longing should follow its prior display of joyshared with the user This requires critically formulating in-tentions behind communicative moves [44] of any affectivebot There are several paths for developing emotional dis-plays To list a few (1) offer one type of consistent emotionalexpressions as Vincent did (2) design emotional expressionsthat may be appropriate for different target groups in tan-dem with the implication below and (3) give users controlover how their chatbots ldquofeelrdquo towards them The caveat forthe third recommendation is that the user control over achatbotrsquos emotions may not aid a chatbotrsquos narrative andit also may not be helpful for all users the associated riskis that user-controlled emotions can render a chatbot lessrelatable More specifically the illusion that a chatbot canauthentically care for or be cared by another being requiressome level of perceived independence in how it ldquofeelsrdquo Werecommend designers to engage with the growing field ofaffective computing [41 45] and its discussion on ethics [8]If a chatbotrsquos goal is bettering usersrsquo psychological statesdesigners must ask if an affective bot delivers the intendedtreatment and what ethical boundaries there are in its dis-plays and elicitation of emotions Designers and users couldcontrol a chatbotrsquos emotions but what emotions a chatbotcan elicit in users is not always a priori foreseeable

Tailor chatbots to different target groups Evenwith one con-struct self-compassion we see a variety of ways a chatbotcan be configured To start people with low self-compassionmay benefit the most from Vincent as our exploratory anal-ysis shows This can mean more compassion focused sce-narios rather than neutral scenarios Women are noted toscore lower on self-compassion [51] yet older women experi-ence greater compassion than older men [29] Chatbots thatconsider gender age andor occupation can be potentiallyhelpful for increasing self-compassion To list a few examplesfor reincarnating Vincent a chatbot could be gendered asfemale or non-binary present a proactive version of compas-sion specified for women (see eg Neff [35] on speaking up

and protecting oneself from harm) talk about exam stresswith students or refer to stressful meetings or workplacebullying with employed individuals Rather than assumingthat one-size-fits-all or extreme personalization will workwe suggest designers to first approach targeted groups toclearly understand their needs For instance whether a self-compassion chatbot for all women is as effective or moreeffective than a more targeted chatbot eg at various levelsof intersectionality like race culture age etc should beconsidered given the time and resources that may be avail-able We recommend that based on research uncoveringpossible ways to design a chatbot that suits different needsand wants should be prioritized

Future works and limitationsOur study opened up new questions to be explored An av-enue to investigate is how a chatbotrsquos use of emotional lan-guage influences its interactants We posit that a sufferingchatbot induces less empathic distress than a suffering hu-man and whether or not this is the case needs to be furtherinvestigated especially for chatbots intended to be therapeu-tic helpers An awareness of onersquos own and othersrsquo sufferingwithout overwhelming empathic distress is suggested to bepossible through compassion [18 47] Hence disambiguat-ing compassion from empathic distress is critical in deploy-ing self-compassion chatbots as instantiations of positivecomputing [6] Different configurations of Vincent basedon peoplersquos gender age or occupation could improve theirself-compassion scores more effectively and if and in whatways this holds true warrants further research

There are limitations to consider Our effect size was lowerthan findings from the Woebot study (d = 044) which had34 participants who ldquoself-identified as experiencing symp-toms of depression and anxietyrdquo [15 p 2] and they measuredsymptoms of depression with the PHQ-9 questionnaire notself-compassion Falconer et alrsquos results on self-compassionscores after embodied VR experience also had a higher ef-fect size with the partial eta-squared of 036 (d = 1487) [14]which was based on 15 participants with depression Weworked with a general non-clinical sample and CG Vincentshowed an effect size of d = 007 (N = 34) and CR Vincentrsquoseffect size was d = 02 (N =33) Follow-up studies on self-compassion chatbots can utilize a larger sample and a moregrounded effect size One explanation for the difference ineffect size is that we did not recruit people who were clini-cally or self-proclaimed to be depressed based on the viewthat preventative mental health care can build resilience forpeople in general While Vincent andWoebot [15] share com-monalities the main measurements and targeted populationdiffered And while self-compassion was the measurementfor us and Falconer et al [14] the technology used sample

size and targeted population differed The gain andor main-tenance of healthy self-compassion as pre-emptive care maynot result in a similarly high effect size but can be psycholog-ically beneficial nonetheless More research is necessary tounderstand long-term consequences of a priori preventativecare vs a posteriori treatment of mental health woes

More broadly peoplersquos engagement with Vincent may re-flect both socially desirable reactions such as politeness to-wards machines as social actors [16 31] as well as emotionalempathy ie the ability to ldquofeel forrdquo VincentWe have not yetconcretely looked into other potential contributing factorsin bringing about self-compassion through human-chatbotinteraction Also what is difficult to gauge is the magnitudeof a chatbotrsquos perceived social and emotional complexitybased solely on messaging or text-based conversations Vin-cent lacked embodied communication which means it didnot use non-verbal modalities such as gaze voice or ges-tures that are critical in various social interactions Vincentwas a uni-modal technological entity that can be extendedthrough other complex emotional displays Thus we havenot established how people would engage with other formsof technology like robots with varying degrees and types ofembodiment alongside different combinations of modalitiesUtilizing technology appropriately for mental health carerequires many comparative renditions

5 CONCLUSIONCompassion is a key moral emotion [19] or motivation [6]that deserves to be further explored through positive comput-ing or technology for well-being Self-compassion can helppeoplersquos overall well-being [52] through kindness towardsoneself connectedness to greater humanity andmindfulnessWhile a chatbot is not a panacea for curing psychologicaldifficulties and is not meant to replace professional helpwe demonstrated that caring for a chatbot can help peoplegain greater self-compassion than being cared for by a chat-bot Our quantitative and qualitative analyses suggest thathuman-chatbot interaction is a promising arena for positivecomputing

REFERENCES[1] Anxiety and Depression Association of America (ADAA) 2016 ADAA

Reviewed Mental Health Apps httpsadaaorgmental-health-appsAccessed 2018-06-25

[2] Arthur Aron Elaine N Aron and Danny Smollan 1992 Inclusion ofOther in the Self Scale and the structure of interpersonal closenessJournal of Personality and Social Psychology 63 4 (1992) 596 httpsdoiorg1010370022-3514634596

[3] Virginia Braun and Victoria Clarke 2006 Using thematic analysisin psychology Qualitative Research in Psychology 3 2 (2006) 77ndash101httpsdoiorg1011911478088706qp063oa

[4] Scott Brave Clifford Nass and Kevin Hutchinson 2005 Computersthat care investigating the effects of orientation of emotion exhibited

practice resulted in a relatively even gender distribution forboth conditions (CR M = 18 F = 14 undisclosed = 1 CG M= 18 F = 15 undisclosed = 1)

Chatbot implementation

Figure 1 Care-receiving Vincent

Vincent2 was built with Googlersquos Dialog flow that wasintegrated to Facebook Messenger3 We purposefully did notvisually design Vincent (the profile icon showed a ldquoVrdquo) todrive Vincentrsquos personality by what was said rather thanhow Vincent looked Participantsrsquo responses did not changeVincentrsquos reactions The usage of limited pre-set responses(Figure 1) was to continue Vincentrsquos narrative whilst allow-ing participants a choice between relevant responses (Woe-bot also interacted with its users in a similar fashion [15])This allowed for greater comparability between participantsin each condition

We had eight scenarios each for CG or CR Vincent and 6neutral scenarios (total 22 scenarios) The neutral scenarioswere the same for both Vincents and aimed for adherenceto daily touch-points in an entertaining way For exampleone neutral scenario is about world records ldquo[] I knew you

2Vincentrsquos Facebook page - httpswwwfacebookcomvincentthebot3Dialog Flow Facebook Integration - httpsdialogflowcomdocsintegrationsfacebook

were an interesting bunch of carbon-based beings but appar-ently you have this thing called lsquoworld recordsrsquo [] [] whatwould you like to be in the world records book forrdquo BothVincents used images and emojis because visual icons arewidespread in digital messaging to express emotions [43]Weused appropriate punctuation marks and positively and neg-atively valenced syntax in accordance to previous researchon text-based emotion expression and detection [20] Mostinputs were close-ended (Figure 1) but we allowed open-ended free inputs at least once per interaction We aimedfor a meaningful comparison in two ways In each condi-tion participantsrsquo interactions were designed to be the samewith limited set of possible reactions to Vincent Betweenthe two Vincents we wanted to clearly distinguish betweenthe recipient and the giver of care Below are excerpts fromCG and CR Vincents (all scenarios are in the supplementarymaterial)

CG Vincent [] see if you can think of a kinder more caringway to motivate yourself to make a change if needed Whatis the most supportive message you can think of thatrsquos in linewith your underlying wish to be healthy and happy Try towrite it belowUser [free input]

CR Vincent What do you think am I the dumbest bot yoursquoveever seen or whatU [free input]V Am I being too hard on myselfU [free input]

CG Vincent was modeled after Woebot [15] supplementedby self-compassion exercises [34] CR Vincent was based onthe Self-Compassion and Self-Criticism scale [13 14] Thescale has scenarios like job rejections unpaid bill remindersand being late to a meeting which were converted to fita chatbot (Figure 1) The eight items of the scale were in-terweaved for conversational storytelling By doing so wejuxtaposed issues that students (the majority of our sample)can face eg distress about failing an exam that Vincentunderwent CR Vincent narrated its story over time by ad-mitting its mistakes feeling inadequate compared to otherchatbots confessing a former lie and asking for confiden-tiality in sharing a non-trivial worryI got a reminder from the server that hosts me Itrsquos like my

house so to say [] I forgot to pay the server fee on time[][] It wouldrsquove taken me only 0004 seconds to make the trans-action you know since Irsquom a robot and all [] this never seemsto happen to the other chatbots on my serverVincent brought up this scenario again later with new

informationRemember our talk a couple of days ago About me forgetting

to pay my server fee in time [] I kind of lied to you [] Ididnrsquot tell you this before because I was a little embarrassedabout it Can you promise me that this stays between us []Irsquove been applying for different jobs and just today I receivedmy third rejection email already The reason that I couldnrsquotpay my bill was because Irsquom running out of money And if Idonrsquot find a job soon Irsquoll get kicked off my server

In sum CG Vincent guided participants through activitieswhile CR Vincent opened up about its everyday mistakesor worries to get support from participants CG Vincentsought to trigger self-compassion in users themselves and CRVincent gave participants opportunities to be compassionatetowards a chatbot

MeasurementsWe collected basic demographic information ie genderage and previous experience with a chatbot at the startOur main measurement was the self-compassion scale on afive point scale [32] with six sub-components (self-kindnessvs self-judgment common humanity vs isolation mind-fulness vs over-identification) which was deployed twicebefore and after the experiment We added a measurementfor opinions about the agent on a seven point scale [4] todetect irregularities between how CG and CR Vincents maybe perceived Our final scale was the Inclusion of Other inthe Self (IOS) Scale a single item on a seven point scale [2]to check howmuch participants identified with Vincent post-hoc We also kept track of two additional aspects to gaugeengagement quality One is the error rate ie the number oftimes Dialogflow crashed during the interaction sometimesrequiring an experimenter to momentarily ldquowizardrdquo [9] torestart the interaction4 The other is the total word countper participant on open ended answers

ProcedureWe built Vincent and wrote our initial scenarios to be testedOur pilot study of three days was with voluntary participants(N = 12) personally recruited by experimenters We checkedif scenario categories (caregiving care-receiving and neu-tral) were clear by asking participant to guess Vincentrsquosintentions and goals per scenario Based on this we onlyadapted neutral scenarios Then we recruited participantsfor the actual experiment Our email invitation was sent outto the TU Eindhoven participant database and interestedparticipants joined the experiment if they used Facebookand could freely express themselves in English The emailcontained a link to Vincentrsquos Facebook page that partici-pants had to go to for a guided tutorial on the experimentpayment information and the informed consent form This

4Restarts happened 37 out of 938 interactions (14 days 67 participants) or394 of the time

form noted that experimenters will look at participantsrsquo dataand that third party technology providers Vincent relied onie Facebook and Google have access to the informationWe added that participantsrsquo personally identifiable informa-tion will not be shared for publication purposes and thattheir voluntary participation means that they can drop outof the study at any point We also stated that Vincent is nota therapist and that they should seek professional help if anypsychological issues are experienced though we targeted anon-clinical population

After the guided tutorial participants filled in the first setof questions on demographic information and self-compassionThen they were assigned to either CG or CR Vincent Theyall began the experiment on the same day For two weeksVincent greeted and sent a password daily and participantshad to repeat the password to Vincent to start the daily inter-action eg Hey me again ) Tell me tHup to start our littletalk After two weeks participants filled in the final surveyon self-compassion the IOS scale opinion on the chatbotdetails for compensation as well as additional comments orfeedback they were willing to share Participants were thenpaid through bank transfer

4 RESULTSWe first present our quantitative analysis and then move onto how people talked with Vincent the qualitative angle Allrelevant tables and graphs as well as CG and CR Vincentsrsquoscenarios are in the supplementary material

Quantitative analysisBefore we forged ahead with hypotheses testing we lookedinto engagement levels of all participants to detect outliersWe had three outliers participants who had less than 20minutes of total interaction time with Vincent Only readingwhat Vincent sent not including giving a response shouldtake one to two minutes per day We expected a minimum of20 to 28 minutes of interaction for two weeks (for all partici-pants the average total time was 36 minutes SD = 10) Ouroutliers spent in total 15 18 and 19 minutes each the threelowest total interaction time Correspondingly their totalword count to open responses reflected low engagement at27 22 and 27 words for the total duration of the experimentthe three lowest total word count out of all participants

We conducted two one-tailed dependent t-tests [7] to an-swer our hypotheses (we set p at 005 with the confidenceinterval of 95) CG Vincent did not result in significantchange (t(31) = -0572 p = 0286 Cohenrsquos d = 007) whencomparing before (M = 3135 SD = 0630) and after (M =3180 SD = 0628) the two weeks but the direction detectedis positive (normality assumed for prior (W = 0987 p = 0978)and post (W = 0989 p = 0982) scores) CR Vincent did showa significant difference (t(31)= -197 p = 0029 Cohenrsquos d

= 02) between prior (M = 3137 SD = 0613) and post (M= 3257 SD = 0558) scores for self-compassion (normalityassumed for prior (W = 0979 p = 0764) and post (W = 0946p = 0114) scores)We conducted exploratory analyses to better understand

our data Through a repeated measures ANOVA we checkedfor the effect of time prior and post self-compassion scores(F(1 62) = 2768 p = 0101 η2 = 0043) Then we checked forthe effect of condition ie CG or CR (F(1 62) = 075 p = 0785η2 = 0001) and the interaction between time and condition(F(1 62) = 0580 p = 0449 η2 = 0009) None were significantWe additionally checked for time condition timeconditioneffects on the three components of self-compassion self-kindness common humanity and mindfulness Only theeffect of time on common humanity was significant F(1 62)= 6059 p = 0017) The table for all statistical tests is in thesupplementary materialWe further dissected our data by gender because previ-

ous research showed that women may score lower on self-compassion [51] Indeed female participants had lower self-compassion scores (M = 305 SD = 013) than men (M = 326SD = 009) at the start but not significantly so (t (4935) =113 p = 026 d = 029) according an independent unequalvariance t-test (normality assumed) We then compared postand prior scores for men and women Menrsquos self compassionscores increased only by 002 as a difference in means andshowed no significant increase (t (33) = -025 p = 040 d =004) However womenrsquos scores showed a significant differ-ence (t (27)= -206 p = 002) between 305 (SD = 071) as thestarting score and 319 (SD = 065) as a posterior score forself-compassion When we scrutinized the gender differencebetween CG and CR Vincents we noticed a more dramaticdifference Women with CR Vincent showed a highly signifi-cant change (t(13) = -289 p = 0006 d = 077) compared towomen with CG Vincent (t(13) = -033 p = 037 d = 009)and both were normally distributed

We wanted to check if mainly gender was at stake or if itwas simply a difference between low vs high scorers on priorself-compassion levels We thus divided all participants intotwo groups based on the average self-compassion score at thestart 314 Those who scored above this were high scorers(M = 371 SD = 037 N = 28) those who scored below werelow scorers (M = 269 SD = 032 N = 36) We included oneparticipant with the average score in the high-scoring groupand this had no impact on significance reached Low scorersgreatly increased their self-compassion scores in terms ofsignificance (t(35) = -341 p = 00008 d = 057) but highscorers did not show improvements (t(27) = 110 p = 086 d= 018) Yet normality was not assumed for both low-scorers(W = 093 p = 003) and high-scorers (W = 091 p = 002)since we divided a normally distributed group into two Thuswe performed a two-sample Wilcoxon rank-sum test to see

if there was a significant difference between low scorers andhigh scorers which was the case at z = 286 and p = 0004 CRVincent improved low scorersrsquo self-compassion significantly(t(17) = -320 p = 0003 d = 075) compared to a marginalsignificance for CG Vincent (t(17) = -175 p = 005 d = 041)There were normal distributions observed for prior and postself-compassion levels of low and high scorersA potential explanation for why low scorers improved

more than high-scorers is regression to the mean Howeverin published literature the average self-compassion score isbetween 25 and 35 [33] and our low scorers have a prior av-erage self-compassion of 269 If regression to the mean is anexplanation we would also expect high-scorers to end witha lower mean yet this is not the case Our high-scorers hadan average prior score of 371 (above average [33]) and theirscores did not decrease after the experiment This may be aceiling effect The low-scorersrsquo improvement is still there itis a highly significant effect even with the Bonferroni correc-tion for all tests (p = 00008) with the post-hoc power of 097The data supports that Vincent enhanced self-compassionfor low-scorersWe had two additional scales one to check if people per-

ceived CG and CR Vincent in a relatively similar way [4]and the other to see how much participants identified withVincent [2] The survey on participantsrsquo opinion of the agentincluded caring likability trustworthiness intelligence dom-inance and submissiveness as items [4] about Vincent BothCG and CR Vincents were perceived to be fairly analogoussave for dominance (α = 048 CG Vincent M = 2677 SD =0794 CR Vincent M = 2656 SD = 0700) and submissiveness(α = 024 CG Vincent M = 2448 SD = 0559 CR Vincent M= 2396 SD = 0636) though none showed a significant dif-ference between two conditions The the Inclusion of Otherin the Self (IOS) scale [2] displayed that participants moreclosely related to CR Vincent (M = 348 SD = 148) thanCG Vincent (M = 306 SD = 139) but not significantly so (t= -115 p = 025 d = 029) Both conditions were normallydistributed (CG W = 096 p = 034 CR W = 097 p = 064)

To summarize our hypothesis that CG Vincent increasesself-compassion was not supported (p = 0286) but the hy-pothesis that CR Vincent increases self-compassion was sup-ported (p = 0029) Our exploratory analyses captured threeunderlying influences on this finding First our ANOVA testsrevealed that the only significant aspect was time as an inde-pendent variable affecting common humanity one elementof self-compassion (p = 0017) Second gender may be a con-tributing factor with women demonstrating a significantincrease in self-compassion (p = 002) for both conditionscombined but not men (p = 040) To add only CR Vincentdemonstrated a highly significant change for women (p =0006) unlike women who interacted with CG Vincent (p =037) Third regardless of gender those who started out with

a low self-compassion score exhibited the most significantchange (p = 00008) for both conditions together Low-scorersmore significantly improvedwith CR Vincent (p = 0003) thanwith CG Vincent (p = 005)

Put together CR Vincent more effectively increased self-compassion than CG Vincent most likely through a signif-icant change in participantsrsquo sense of common humanitymore so than self-kindness and mindfulness (as our ANOVAtest showed) Finding common humanity can be inclusive ofchatbots Women specifically those with CR Vincent weresignificantly more affected than men However low-scorersof both genders benefited the most compared to high-scorersespecially those with CR Vincent CG and CR Vincents werenot perceived to be significantly different except for a lowersimilarity regarding the dominance-submissive trait Par-ticipants in the CR condition may have felt that Vincent ismore like them (ISO scale) though this difference was notsignificant

Qualitative analysisFor our qualitative analysis we used the corpus of free re-sponses that participants had typed during their interactionswith Vincent We will first present a descriptive analysisof the interactions that people had with CG and CR Vin-cents followed by our thematic analysis [5] and interpretiveanalysis [48] Participantsrsquo responses to CR Vincent were onaverage 9353 words (SD = 4796) and to CG Vincent 11247words (SD = 6399) for two weeks While our data set is notabundant in terms of word count we believe a qualitativelook at how participants interacted with Vincent is valuable

Descriptive analysis CGVincent guided participants throughself-compassion exercises [34] that did not require them toactively voice aspects of self-compassion they mainly hadto read about it This resulted in fewer instances of self-compassion themes in CG Vincent since they only occurredwhen Vincent asked participants to comfort themselves Toadd participantsrsquo willingness to engage with CG Vincentrsquosprobes differed It asked specific questions or provided ashort task for free input eg ldquowrite down a difficulty youhaverdquo Many answers to this were short ldquowake up earlyrdquoor ldquoI often feel alonerdquo Some participants opened up moreldquoI have a difficulty in expressing myself when I am underdifficult situationsrdquo or ldquoI am studying abroad far from homeand family and friends Different culture language edu-cational standardrdquo CG Vincent asked a follow-up questionldquohow did it make you feelrdquo we again got simple answerslike ldquogoodrdquo or ldquonormalrdquo or longer expressions ldquoI feel morerefreshedrdquo or ldquonot really better or worserdquo In other instancesCG Vincent allowed participants to give themselves simpleself-assurance ldquoI can do itrdquo plan for long-term goals ldquoonceI have graduated I can schedule a good routine to target a

fitter and healthier lifestyle Just hang in there a little longerrdquoor dig deeper ldquopeople around me would be happier if I washappierrdquo Thus CG Vincent provided a daily touch point forself-reflection admittance of everyday suffering ldquopep-talkrdquoor deeper self-insight which may or may not directly relateto self-compassion for all participants

In contrast CR Vincent frequently asked for help and con-sequently received many self-compassion related answersThe narrative was the focus for CR Vincent It was able tobecome more vulnerable over time by admitting its own ev-eryday hardships as a chatbot which led it to seek opinion oradvice For example CR Vincent asked ldquowhat do you thinkam I the dumbest bot yoursquove ever seen or what Am I beingtoo hard on myselfrdquo To this participants responded in dif-ferent ways ldquoI think yoursquore the funniest bot that Irsquove everseen mdash yes you are in some situationsrdquo ldquoNo but you shouldexpect (that) a bot is very smartly programmed and know allmdash Maybe I do not knowrdquo or ldquothe world needs bots like youAnd itrsquos usual to get rejected sometimes just keep on goingand yoursquoll find a job soon enoughrdquo However CR Vincentrsquoscries for help did not always result in necessarily compassion-ate replies Many users stuck to pragmatic answers relatedto the topic of the problem Even though all of CR Vincentrsquosscenarios were intended to generate compassion towardsVincent pragmatic replies indicate that not everyone willdemonstrate compassionate responses in every instanceThe difference between CG and CR Vincents is that be-

ing compassionate towards another being in a conversa-tional narrative context is unlike doing guided exerciseson self-compassion about oneself The frequency of con-structing compassionate replies is a way to practice self-compassion users of CR Vincent spent more time practicingself-compassion than those with CG Vincent Therefore CRVincent was more effective than CG Vincent since CR Vin-cent provided more opportunities to be compassionate Thecaveat is that the link between frequency of practice andincrease in self-compassion may not be direct Althoughmindfulness and self-kindness were most often observedonly common humanity improved significantly accordingto our exploratory quantitative analysis Finding commonhumanity in and through a chatbot is also a strong theme inour thematic analysis

Thematic analysis We categorized our data according tothree pillars of self-compassion [32] as displayed in Table 1While all three sub-components were present in both care-receiving and caregiving conditions more instances occurredwith CR Vincent The numbers below a theme (Table 1) arecounts of how many times it occurred in each condition Allquotes below were to CR VincentAs quotes in Table 1 suggest many participants offered

helpful advice to CR Vincent Vincent showed appreciation

Theme QuoteMindfulnessCaregiving 3Care-receiving 25

There are worse things thatcould happen What has hap-pened has happened

Self-kindnessCaregiving 7Care-receiving 21

Go do something fun todaylike watching a movie Staypositive and keep trying untilyou succeed

Common humanityCaregiving 0Care-receiving 11

Everyone makes mistakesJust remember that it can hap-pen to anyone and that itrsquos notyour fault

Table 1 Self-compassion sub-components

with follow-up statements like ldquoyou always make me feel bet-terrdquo Negative counterparts to three pillars of self-compassionwere not strongly present ie self-judgment was detectedfour times for CR Vincent and once for CG Vincent isolationwas noted once for CR Vincent but none for CG Vincentand over-identification was neither present for CG nor CRVincent

For both conditions people were mostly friendly to Vin-cent and there were no swear words or abusive languagedisplayed The most hostile comment was ldquoyoursquove beenpretty dumbrdquo to CR Vincent and we encountered suchldquoput-downsrdquo only twice There were additional topics thatemerged through open thematic analysis They are sum-marized in Table 2 and these themes could also pertain toself-compassion themes (messages to CGVincent are markedwith ldquoCGrdquo and otherwise they were to CR Vincent)

Theme QuotePragmatismCaregiving 0Care-receiving 41

Maybe next time make a bet-ter planning and make sureyoursquove got enough time )

Perspective-takingCaregiving 0Care-receiving 10

I would find a window toclimb in But maybe in yourcase better try to hack into thefolder [] be proud of thebot that you are

Engagement vsDistantiationCaregiving 27 vs 6Care-receiving 5 vs 11

A girl told me she loves meAnd I love her too (CG) vsSorry itrsquos confidential (CG)

Positive vs NegativeCaregiving 74 vs 9Care-receiving 5 vs 2

I was laying in a field full offlowers trying out my newukelele (CG) vs I hate pink

Table 2 Free-input themes

People interacted with CG and CR Vincents differently(Table 2) Giving pragmatic advice to Vincent and taking itsperspective as a chatbot were themes only found in the CRcondition Rather than tending to Vincent by giving emo-tional support participants gave practical advice on what todo better Examples of perspective-taking are recommendingVincent to ldquohack into the folderrdquo or to use ldquobrute forcerdquo tech-niques to gain access participants thought like a chatbot tohelp a chatbotSome participants revealed more personal information

to CG Vincent (theme engagement) and took interest inVincent by asking questions back eg ldquowhat (did you) dofor money before nowrdquo or writing lengthy responses Someshared information was very intimate in nature eg ldquoIrsquomgoing to kiss the girl next to me in 5 secondsrdquo Since CGVincent asked participants to write about themselves thisskewed engagement (the amount of textual response) to-wards CG Vincent Participants distanced themselves fromVincent only a few times by stating that certain informationwas confidential or not showing interest in getting to knowVincent eg ldquosorry I do not know what interests yourdquo Thelast theme on positive vs negative attitude was primarilypresent in the CG condition this theme was mostly aboutattitudes participants had about themselves and their livesnot about Vincent Most participants shared positive lifeevents eg getting an internship cooking something deli-cious Though negative attitudes were minimal they rangedfrom more mundane states eg feeling ldquoawkwardrdquo to moredramatic states eg ldquonot die within 2 weeksrdquoTo summarize Tables 1 and 2 self-compassion sub- com-

ponents were more present with CR Vincent suggestingthat giving compassion to Vincent (or another being) thantowards oneself may be more natural in conversational con-texts And mindfulness most frequently occurred (Table 1)As for emergent themes in Table 2 participants gave prag-matic advice to CR Vincent and often practiced perspective-taking Yet CG Vincent allowed for more self-expression ifparticipants were open to communicate as shown by greaterinstances of engagement and positive remarks about every-day situations In a few instances we detected deeply per-sonal messages on ups and downs of relationships and self-deprecating thoughts Mostly participants shared positivedaily news with CG Vincent and helpful or uplifting remarkswith CR Vincent

Interpretive analysis We now offer a broader interpreta-tion of our data by incorporating participantsrsquo open-endedresponses to an item on the final survey The main theme isbonding between participants and Vincent though not allbonded with Vincent in the same way To explain this weprovide three subthemes that underlie the bonding processwith Vincent Our primary focus was on CR Vincent

Relatability leads to believability Participantsrsquo ability toextend the sense of common humanity to a chatbot touchesupon anthropomorphism CR Vincent was comforted as ifit were a human eg ldquoitrsquos human to make mistakesrdquo (CR)while its problems were addressed to its ldquochatbot worldrdquo egldquocommunicate whatrsquos going on to your fellow chatbotsrdquo (CR)For one participant even Vincentrsquos limitation of having astrict script was anthropomorphized ie ldquoVincent is like thefriend who always speaks about himself and what he haslearned or done and sometimes out of courtesy (not out ofcuriosity) asks how you are doing - but doesnrsquot listen to youranswer or what you actually have to say he just goes onwith his own thingrdquo (CG) Such attributed anthropomorphictraits depended on participantsrsquo willingness take Vincentrsquosperspective as a chatbotCR Vincentrsquos blunders were based on common human

mishaps like being late for a meeting and dealing with un-paid bills (scenarios from [13]) Yet none of our participantsquestioned whether or not a chatbot had meetings to attendor bills to pay Vincentrsquos narrative was on how a chatbotcould be late (new updates took longer than expected) orhow it could have bills (Vincent needs to pay the hostingserver) and our participants went along with imagined sce-narios Vincent faced Rather than questioning the parametersof our scenarios on realism participants thought of how tosolve Vincentrsquos problems within the parameters of a chat-botrsquos world When relevant CR Vincent played up the ironyof having human struggles as a chatbot eg ldquoall I am is lit-erally a piece of code and I failed a programming courserdquoVincent became believable because its struggles were relat-able Granting Vincent human-likness was less literal in howpeople bonded with Vincent Vincent did not try to appearhuman but it socialized with participants about its strugglesthat humans also had People related to Vincentrsquos strugglesand believed that such struggles could arise for chatbotsShared history can lead to attachment Conversations be-

tween people as well as in human-computer interactionbecome shared history over time For instance ldquo[] commu-nicating with Vincent everyday for two weeks builds somekind of habit It makes me notice its presence and absence(which might be good) I think it has a potential to be a goodcompanion and improve the mood especially if someone isfeeling lonelyrdquo (CG) Thus frequent communication with achatbot in a given duration can form expectations ldquoI reallymissed Vincent when he started our conversation laterdquo (CR)The level of attachment for some participants was higherthan others eg after the experiment we saw reactions suchas ldquocan I keep himrdquo (CG)When Vincent prepared participants for its daily good-

byes eg ldquoI have some chatbot things to do Defragmentmy server stack Buy aluminum foil to make fashionablehats withrdquo what was intended as humor can be interpreted

differently ie server defragmentation could be life-or-deathfor a chatbot Some people can be confused worried or evenangered when a chatbot they care about does not respondThus one reaction was ldquothe asshole decided to delete itsstack and when I said itrsquod die it just didnrsquot reply You canrsquotgo making people worried about a freaking chatbotrdquo (CR)People may miss a chatbot that suddenly leaves them orsincerely worry about its well-being This is the positiveand negative aspect of a relatable a chatbot some partic-ipants found common-humanity in Vincent and of thoseparticipants a few possibly related more through empathicdistress rather than through compassion If two weeks canbring about strong signs of attachment longer periods ofinteraction may heighten the level of attachment to differentdegrees and in different ways per person

Emotional reciprocity with chatbots As mentioned beforemost participants were able to respond to CR Vincentrsquos emo-tional displays on a practical level eg recommending howto fix a problem or advising Vincent on how to adjust itsemotions eg telling Vincent to stay positive To add somepeople may not appreciate chatbots demonstrating feelingsOthers may reciprocate or feel comforted by a chatbotrsquos ex-pressed emotions even if a chatbot is perceived as incapableof having emotions The more nuanced point is that Vin-centrsquos display of emotions was noted to bring conflictingfeelings For instance ldquowhen Vincent would show emotions(for example lsquolove talking to yoursquo lsquomiss yoursquo) that would feelweird because I know I am talking to a chatbot and it proba-bly is not that developed that it does have feelings But theusage of such words does feels nice compared to when a hu-man would say them So I had conflicted feelings about thesekind of expressionsrdquo (CG) The participant felt conflictedabout how to process Vincentrsquos emotional outreach

Importantly the participant suggested heshemay bemorecomfortable with a chatbot saying ldquomiss yourdquo than a humanTo conjecture the participant could mean that there was nosocial pressure due to a chatbot not expecting or needinghimher to say ldquoI miss you toordquo People often feel obligatedto respond to sincere emotions of others with similarly va-lenced emotional displays even if they do not feel the samesincere emotions towards them Such pressure may not holdfor technological entities Perhaps to miss someone impliesa certain history in a relationship so to hear that from aperson one met less than two weeks ago may feel awkwardor insincere whereas a chatbot would not be expected toknow or abide by certain social conventions If two peopleknew beforehand they will get to know each other for a maxi-mum duration of two weeks (as our participants knew beforemeeting Vincent) and never be in touch again their emo-tional performance may adjust accordingly The timescalefor intensifying socially acceptable emotional expressions

in human-chatbot interactions and human-human interac-tions may differ The ldquolifespanrdquo of a chatbot is not equatableto a personrsquos lifespan And the distinction between super-ficial vs genuine emotional displays from and to a chatbotis not entirely equatable to emotions people share and re-ciprocate between each other Currently we do not haveestablished norms on how emotions between humans andbots areshould be managed We suggest there may be dis-tinct differences compared to emotions in human-humanrelationships

Discussion and design implicationsCR Vincent adds depth to the CASA paradigm [16 31]mdash notonly do people treat a chatbot as an anthropomorphizedsocial agent but they themselves are affected by a chatbotto the extent that their self-compassion can increase whenthey are compassionate towards a chatbot Brave et alrsquosinsight on embodied conversational agents is that ldquojust aspeople respond to being cared about by other people usersrespond positively to agents that carerdquo [4 p 174] We addthat just as giving care to another human can increase onersquosself-compassion [5] caring for a chatbot can enhance onersquosown self-compassion If the dominant question has beenldquowhat can technology do for usrdquo Vincent demonstrates thatby exploring ldquowhat can we do for technologyrdquo we inad-vertently benefit from technology potentially more so thanwhen we only shape technology to serve us This claim isspecified to bots in the mental health domain and our goalwas to increase self-compassion as a target for well-being[52] rather than to reduce clinically defined symptoms ofpsychological ill-being We present our design implicationsbelow on building chatbots for psychological health carewhich primarily stem from our interpretive analysis Ourimplications are inter-related starting points that should becontextualized for each research and deployment processGive users more closed-inputs or free-input optionsMany

participants felt limited in responses they could give to Vin-cent They wanted to write to Vincent without having anypreset answers or needed more options A recommendationis to use natural language processing for a chatbot whichwill rely less on a pre-planned narrative arc and build moreon what users say This will require a longer developmentperiod The simpler option is to provide users with morefixed responses (three to four) and more opportunities foropen input

Develop a chatbotrsquos story with users Conversational agentscan be powerful storytellers [25 36] even without complexAI To deliver co-storytelling as shared history with interac-tants we suggest designers to create flexible narrative pa-rameters that people can creatively use to relate to a chatbotVincent was able to tell its story but it was less interactive inthat people could follow along with limited reaction options

due to the nature of our experiment There can be additionalcomplexities that designers can add For instance the nar-rative can take a different route depending on which closedinput options users click on We have utilized a limited num-ber of messages called ldquopathsrdquo (supplementary material) thatVincent could give depending on closed input responsesYet this practice did not change Vincentrsquos narrative Givinga chatbot ldquomemoryrdquo be it knowing basic information likenames or a more involved retention of what users say canenhance conversational storytellingTread carefully with emotional expressionsWe suggest a

broader view on what emotions are by considering inter-related emotions that develop over time For example for abot to miss someone assumes a botrsquos happinessenjoymentexperienced during a prior interaction with a user a botrsquosability to feel longing should follow its prior display of joyshared with the user This requires critically formulating in-tentions behind communicative moves [44] of any affectivebot There are several paths for developing emotional dis-plays To list a few (1) offer one type of consistent emotionalexpressions as Vincent did (2) design emotional expressionsthat may be appropriate for different target groups in tan-dem with the implication below and (3) give users controlover how their chatbots ldquofeelrdquo towards them The caveat forthe third recommendation is that the user control over achatbotrsquos emotions may not aid a chatbotrsquos narrative andit also may not be helpful for all users the associated riskis that user-controlled emotions can render a chatbot lessrelatable More specifically the illusion that a chatbot canauthentically care for or be cared by another being requiressome level of perceived independence in how it ldquofeelsrdquo Werecommend designers to engage with the growing field ofaffective computing [41 45] and its discussion on ethics [8]If a chatbotrsquos goal is bettering usersrsquo psychological statesdesigners must ask if an affective bot delivers the intendedtreatment and what ethical boundaries there are in its dis-plays and elicitation of emotions Designers and users couldcontrol a chatbotrsquos emotions but what emotions a chatbotcan elicit in users is not always a priori foreseeable

Tailor chatbots to different target groups Evenwith one con-struct self-compassion we see a variety of ways a chatbotcan be configured To start people with low self-compassionmay benefit the most from Vincent as our exploratory anal-ysis shows This can mean more compassion focused sce-narios rather than neutral scenarios Women are noted toscore lower on self-compassion [51] yet older women experi-ence greater compassion than older men [29] Chatbots thatconsider gender age andor occupation can be potentiallyhelpful for increasing self-compassion To list a few examplesfor reincarnating Vincent a chatbot could be gendered asfemale or non-binary present a proactive version of compas-sion specified for women (see eg Neff [35] on speaking up

and protecting oneself from harm) talk about exam stresswith students or refer to stressful meetings or workplacebullying with employed individuals Rather than assumingthat one-size-fits-all or extreme personalization will workwe suggest designers to first approach targeted groups toclearly understand their needs For instance whether a self-compassion chatbot for all women is as effective or moreeffective than a more targeted chatbot eg at various levelsof intersectionality like race culture age etc should beconsidered given the time and resources that may be avail-able We recommend that based on research uncoveringpossible ways to design a chatbot that suits different needsand wants should be prioritized

Future works and limitationsOur study opened up new questions to be explored An av-enue to investigate is how a chatbotrsquos use of emotional lan-guage influences its interactants We posit that a sufferingchatbot induces less empathic distress than a suffering hu-man and whether or not this is the case needs to be furtherinvestigated especially for chatbots intended to be therapeu-tic helpers An awareness of onersquos own and othersrsquo sufferingwithout overwhelming empathic distress is suggested to bepossible through compassion [18 47] Hence disambiguat-ing compassion from empathic distress is critical in deploy-ing self-compassion chatbots as instantiations of positivecomputing [6] Different configurations of Vincent basedon peoplersquos gender age or occupation could improve theirself-compassion scores more effectively and if and in whatways this holds true warrants further research

There are limitations to consider Our effect size was lowerthan findings from the Woebot study (d = 044) which had34 participants who ldquoself-identified as experiencing symp-toms of depression and anxietyrdquo [15 p 2] and they measuredsymptoms of depression with the PHQ-9 questionnaire notself-compassion Falconer et alrsquos results on self-compassionscores after embodied VR experience also had a higher ef-fect size with the partial eta-squared of 036 (d = 1487) [14]which was based on 15 participants with depression Weworked with a general non-clinical sample and CG Vincentshowed an effect size of d = 007 (N = 34) and CR Vincentrsquoseffect size was d = 02 (N =33) Follow-up studies on self-compassion chatbots can utilize a larger sample and a moregrounded effect size One explanation for the difference ineffect size is that we did not recruit people who were clini-cally or self-proclaimed to be depressed based on the viewthat preventative mental health care can build resilience forpeople in general While Vincent andWoebot [15] share com-monalities the main measurements and targeted populationdiffered And while self-compassion was the measurementfor us and Falconer et al [14] the technology used sample

size and targeted population differed The gain andor main-tenance of healthy self-compassion as pre-emptive care maynot result in a similarly high effect size but can be psycholog-ically beneficial nonetheless More research is necessary tounderstand long-term consequences of a priori preventativecare vs a posteriori treatment of mental health woes

More broadly peoplersquos engagement with Vincent may re-flect both socially desirable reactions such as politeness to-wards machines as social actors [16 31] as well as emotionalempathy ie the ability to ldquofeel forrdquo VincentWe have not yetconcretely looked into other potential contributing factorsin bringing about self-compassion through human-chatbotinteraction Also what is difficult to gauge is the magnitudeof a chatbotrsquos perceived social and emotional complexitybased solely on messaging or text-based conversations Vin-cent lacked embodied communication which means it didnot use non-verbal modalities such as gaze voice or ges-tures that are critical in various social interactions Vincentwas a uni-modal technological entity that can be extendedthrough other complex emotional displays Thus we havenot established how people would engage with other formsof technology like robots with varying degrees and types ofembodiment alongside different combinations of modalitiesUtilizing technology appropriately for mental health carerequires many comparative renditions

5 CONCLUSIONCompassion is a key moral emotion [19] or motivation [6]that deserves to be further explored through positive comput-ing or technology for well-being Self-compassion can helppeoplersquos overall well-being [52] through kindness towardsoneself connectedness to greater humanity andmindfulnessWhile a chatbot is not a panacea for curing psychologicaldifficulties and is not meant to replace professional helpwe demonstrated that caring for a chatbot can help peoplegain greater self-compassion than being cared for by a chat-bot Our quantitative and qualitative analyses suggest thathuman-chatbot interaction is a promising arena for positivecomputing

REFERENCES[1] Anxiety and Depression Association of America (ADAA) 2016 ADAA

Reviewed Mental Health Apps httpsadaaorgmental-health-appsAccessed 2018-06-25

[2] Arthur Aron Elaine N Aron and Danny Smollan 1992 Inclusion ofOther in the Self Scale and the structure of interpersonal closenessJournal of Personality and Social Psychology 63 4 (1992) 596 httpsdoiorg1010370022-3514634596

[3] Virginia Braun and Victoria Clarke 2006 Using thematic analysisin psychology Qualitative Research in Psychology 3 2 (2006) 77ndash101httpsdoiorg1011911478088706qp063oa

[4] Scott Brave Clifford Nass and Kevin Hutchinson 2005 Computersthat care investigating the effects of orientation of emotion exhibited

to pay my server fee in time [] I kind of lied to you [] Ididnrsquot tell you this before because I was a little embarrassedabout it Can you promise me that this stays between us []Irsquove been applying for different jobs and just today I receivedmy third rejection email already The reason that I couldnrsquotpay my bill was because Irsquom running out of money And if Idonrsquot find a job soon Irsquoll get kicked off my server

In sum CG Vincent guided participants through activitieswhile CR Vincent opened up about its everyday mistakesor worries to get support from participants CG Vincentsought to trigger self-compassion in users themselves and CRVincent gave participants opportunities to be compassionatetowards a chatbot

MeasurementsWe collected basic demographic information ie genderage and previous experience with a chatbot at the startOur main measurement was the self-compassion scale on afive point scale [32] with six sub-components (self-kindnessvs self-judgment common humanity vs isolation mind-fulness vs over-identification) which was deployed twicebefore and after the experiment We added a measurementfor opinions about the agent on a seven point scale [4] todetect irregularities between how CG and CR Vincents maybe perceived Our final scale was the Inclusion of Other inthe Self (IOS) Scale a single item on a seven point scale [2]to check howmuch participants identified with Vincent post-hoc We also kept track of two additional aspects to gaugeengagement quality One is the error rate ie the number oftimes Dialogflow crashed during the interaction sometimesrequiring an experimenter to momentarily ldquowizardrdquo [9] torestart the interaction4 The other is the total word countper participant on open ended answers

ProcedureWe built Vincent and wrote our initial scenarios to be testedOur pilot study of three days was with voluntary participants(N = 12) personally recruited by experimenters We checkedif scenario categories (caregiving care-receiving and neu-tral) were clear by asking participant to guess Vincentrsquosintentions and goals per scenario Based on this we onlyadapted neutral scenarios Then we recruited participantsfor the actual experiment Our email invitation was sent outto the TU Eindhoven participant database and interestedparticipants joined the experiment if they used Facebookand could freely express themselves in English The emailcontained a link to Vincentrsquos Facebook page that partici-pants had to go to for a guided tutorial on the experimentpayment information and the informed consent form This

4Restarts happened 37 out of 938 interactions (14 days 67 participants) or394 of the time

form noted that experimenters will look at participantsrsquo dataand that third party technology providers Vincent relied onie Facebook and Google have access to the informationWe added that participantsrsquo personally identifiable informa-tion will not be shared for publication purposes and thattheir voluntary participation means that they can drop outof the study at any point We also stated that Vincent is nota therapist and that they should seek professional help if anypsychological issues are experienced though we targeted anon-clinical population

After the guided tutorial participants filled in the first setof questions on demographic information and self-compassionThen they were assigned to either CG or CR Vincent Theyall began the experiment on the same day For two weeksVincent greeted and sent a password daily and participantshad to repeat the password to Vincent to start the daily inter-action eg Hey me again ) Tell me tHup to start our littletalk After two weeks participants filled in the final surveyon self-compassion the IOS scale opinion on the chatbotdetails for compensation as well as additional comments orfeedback they were willing to share Participants were thenpaid through bank transfer

4 RESULTSWe first present our quantitative analysis and then move onto how people talked with Vincent the qualitative angle Allrelevant tables and graphs as well as CG and CR Vincentsrsquoscenarios are in the supplementary material

Quantitative analysisBefore we forged ahead with hypotheses testing we lookedinto engagement levels of all participants to detect outliersWe had three outliers participants who had less than 20minutes of total interaction time with Vincent Only readingwhat Vincent sent not including giving a response shouldtake one to two minutes per day We expected a minimum of20 to 28 minutes of interaction for two weeks (for all partici-pants the average total time was 36 minutes SD = 10) Ouroutliers spent in total 15 18 and 19 minutes each the threelowest total interaction time Correspondingly their totalword count to open responses reflected low engagement at27 22 and 27 words for the total duration of the experimentthe three lowest total word count out of all participants

We conducted two one-tailed dependent t-tests [7] to an-swer our hypotheses (we set p at 005 with the confidenceinterval of 95) CG Vincent did not result in significantchange (t(31) = -0572 p = 0286 Cohenrsquos d = 007) whencomparing before (M = 3135 SD = 0630) and after (M =3180 SD = 0628) the two weeks but the direction detectedis positive (normality assumed for prior (W = 0987 p = 0978)and post (W = 0989 p = 0982) scores) CR Vincent did showa significant difference (t(31)= -197 p = 0029 Cohenrsquos d

= 02) between prior (M = 3137 SD = 0613) and post (M= 3257 SD = 0558) scores for self-compassion (normalityassumed for prior (W = 0979 p = 0764) and post (W = 0946p = 0114) scores)We conducted exploratory analyses to better understand

our data Through a repeated measures ANOVA we checkedfor the effect of time prior and post self-compassion scores(F(1 62) = 2768 p = 0101 η2 = 0043) Then we checked forthe effect of condition ie CG or CR (F(1 62) = 075 p = 0785η2 = 0001) and the interaction between time and condition(F(1 62) = 0580 p = 0449 η2 = 0009) None were significantWe additionally checked for time condition timeconditioneffects on the three components of self-compassion self-kindness common humanity and mindfulness Only theeffect of time on common humanity was significant F(1 62)= 6059 p = 0017) The table for all statistical tests is in thesupplementary materialWe further dissected our data by gender because previ-

ous research showed that women may score lower on self-compassion [51] Indeed female participants had lower self-compassion scores (M = 305 SD = 013) than men (M = 326SD = 009) at the start but not significantly so (t (4935) =113 p = 026 d = 029) according an independent unequalvariance t-test (normality assumed) We then compared postand prior scores for men and women Menrsquos self compassionscores increased only by 002 as a difference in means andshowed no significant increase (t (33) = -025 p = 040 d =004) However womenrsquos scores showed a significant differ-ence (t (27)= -206 p = 002) between 305 (SD = 071) as thestarting score and 319 (SD = 065) as a posterior score forself-compassion When we scrutinized the gender differencebetween CG and CR Vincents we noticed a more dramaticdifference Women with CR Vincent showed a highly signifi-cant change (t(13) = -289 p = 0006 d = 077) compared towomen with CG Vincent (t(13) = -033 p = 037 d = 009)and both were normally distributed

We wanted to check if mainly gender was at stake or if itwas simply a difference between low vs high scorers on priorself-compassion levels We thus divided all participants intotwo groups based on the average self-compassion score at thestart 314 Those who scored above this were high scorers(M = 371 SD = 037 N = 28) those who scored below werelow scorers (M = 269 SD = 032 N = 36) We included oneparticipant with the average score in the high-scoring groupand this had no impact on significance reached Low scorersgreatly increased their self-compassion scores in terms ofsignificance (t(35) = -341 p = 00008 d = 057) but highscorers did not show improvements (t(27) = 110 p = 086 d= 018) Yet normality was not assumed for both low-scorers(W = 093 p = 003) and high-scorers (W = 091 p = 002)since we divided a normally distributed group into two Thuswe performed a two-sample Wilcoxon rank-sum test to see

if there was a significant difference between low scorers andhigh scorers which was the case at z = 286 and p = 0004 CRVincent improved low scorersrsquo self-compassion significantly(t(17) = -320 p = 0003 d = 075) compared to a marginalsignificance for CG Vincent (t(17) = -175 p = 005 d = 041)There were normal distributions observed for prior and postself-compassion levels of low and high scorersA potential explanation for why low scorers improved

more than high-scorers is regression to the mean Howeverin published literature the average self-compassion score isbetween 25 and 35 [33] and our low scorers have a prior av-erage self-compassion of 269 If regression to the mean is anexplanation we would also expect high-scorers to end witha lower mean yet this is not the case Our high-scorers hadan average prior score of 371 (above average [33]) and theirscores did not decrease after the experiment This may be aceiling effect The low-scorersrsquo improvement is still there itis a highly significant effect even with the Bonferroni correc-tion for all tests (p = 00008) with the post-hoc power of 097The data supports that Vincent enhanced self-compassionfor low-scorersWe had two additional scales one to check if people per-

ceived CG and CR Vincent in a relatively similar way [4]and the other to see how much participants identified withVincent [2] The survey on participantsrsquo opinion of the agentincluded caring likability trustworthiness intelligence dom-inance and submissiveness as items [4] about Vincent BothCG and CR Vincents were perceived to be fairly analogoussave for dominance (α = 048 CG Vincent M = 2677 SD =0794 CR Vincent M = 2656 SD = 0700) and submissiveness(α = 024 CG Vincent M = 2448 SD = 0559 CR Vincent M= 2396 SD = 0636) though none showed a significant dif-ference between two conditions The the Inclusion of Otherin the Self (IOS) scale [2] displayed that participants moreclosely related to CR Vincent (M = 348 SD = 148) thanCG Vincent (M = 306 SD = 139) but not significantly so (t= -115 p = 025 d = 029) Both conditions were normallydistributed (CG W = 096 p = 034 CR W = 097 p = 064)

To summarize our hypothesis that CG Vincent increasesself-compassion was not supported (p = 0286) but the hy-pothesis that CR Vincent increases self-compassion was sup-ported (p = 0029) Our exploratory analyses captured threeunderlying influences on this finding First our ANOVA testsrevealed that the only significant aspect was time as an inde-pendent variable affecting common humanity one elementof self-compassion (p = 0017) Second gender may be a con-tributing factor with women demonstrating a significantincrease in self-compassion (p = 002) for both conditionscombined but not men (p = 040) To add only CR Vincentdemonstrated a highly significant change for women (p =0006) unlike women who interacted with CG Vincent (p =037) Third regardless of gender those who started out with

a low self-compassion score exhibited the most significantchange (p = 00008) for both conditions together Low-scorersmore significantly improvedwith CR Vincent (p = 0003) thanwith CG Vincent (p = 005)

Put together CR Vincent more effectively increased self-compassion than CG Vincent most likely through a signif-icant change in participantsrsquo sense of common humanitymore so than self-kindness and mindfulness (as our ANOVAtest showed) Finding common humanity can be inclusive ofchatbots Women specifically those with CR Vincent weresignificantly more affected than men However low-scorersof both genders benefited the most compared to high-scorersespecially those with CR Vincent CG and CR Vincents werenot perceived to be significantly different except for a lowersimilarity regarding the dominance-submissive trait Par-ticipants in the CR condition may have felt that Vincent ismore like them (ISO scale) though this difference was notsignificant

Qualitative analysisFor our qualitative analysis we used the corpus of free re-sponses that participants had typed during their interactionswith Vincent We will first present a descriptive analysisof the interactions that people had with CG and CR Vin-cents followed by our thematic analysis [5] and interpretiveanalysis [48] Participantsrsquo responses to CR Vincent were onaverage 9353 words (SD = 4796) and to CG Vincent 11247words (SD = 6399) for two weeks While our data set is notabundant in terms of word count we believe a qualitativelook at how participants interacted with Vincent is valuable

Descriptive analysis CGVincent guided participants throughself-compassion exercises [34] that did not require them toactively voice aspects of self-compassion they mainly hadto read about it This resulted in fewer instances of self-compassion themes in CG Vincent since they only occurredwhen Vincent asked participants to comfort themselves Toadd participantsrsquo willingness to engage with CG Vincentrsquosprobes differed It asked specific questions or provided ashort task for free input eg ldquowrite down a difficulty youhaverdquo Many answers to this were short ldquowake up earlyrdquoor ldquoI often feel alonerdquo Some participants opened up moreldquoI have a difficulty in expressing myself when I am underdifficult situationsrdquo or ldquoI am studying abroad far from homeand family and friends Different culture language edu-cational standardrdquo CG Vincent asked a follow-up questionldquohow did it make you feelrdquo we again got simple answerslike ldquogoodrdquo or ldquonormalrdquo or longer expressions ldquoI feel morerefreshedrdquo or ldquonot really better or worserdquo In other instancesCG Vincent allowed participants to give themselves simpleself-assurance ldquoI can do itrdquo plan for long-term goals ldquoonceI have graduated I can schedule a good routine to target a

fitter and healthier lifestyle Just hang in there a little longerrdquoor dig deeper ldquopeople around me would be happier if I washappierrdquo Thus CG Vincent provided a daily touch point forself-reflection admittance of everyday suffering ldquopep-talkrdquoor deeper self-insight which may or may not directly relateto self-compassion for all participants

In contrast CR Vincent frequently asked for help and con-sequently received many self-compassion related answersThe narrative was the focus for CR Vincent It was able tobecome more vulnerable over time by admitting its own ev-eryday hardships as a chatbot which led it to seek opinion oradvice For example CR Vincent asked ldquowhat do you thinkam I the dumbest bot yoursquove ever seen or what Am I beingtoo hard on myselfrdquo To this participants responded in dif-ferent ways ldquoI think yoursquore the funniest bot that Irsquove everseen mdash yes you are in some situationsrdquo ldquoNo but you shouldexpect (that) a bot is very smartly programmed and know allmdash Maybe I do not knowrdquo or ldquothe world needs bots like youAnd itrsquos usual to get rejected sometimes just keep on goingand yoursquoll find a job soon enoughrdquo However CR Vincentrsquoscries for help did not always result in necessarily compassion-ate replies Many users stuck to pragmatic answers relatedto the topic of the problem Even though all of CR Vincentrsquosscenarios were intended to generate compassion towardsVincent pragmatic replies indicate that not everyone willdemonstrate compassionate responses in every instanceThe difference between CG and CR Vincents is that be-

ing compassionate towards another being in a conversa-tional narrative context is unlike doing guided exerciseson self-compassion about oneself The frequency of con-structing compassionate replies is a way to practice self-compassion users of CR Vincent spent more time practicingself-compassion than those with CG Vincent Therefore CRVincent was more effective than CG Vincent since CR Vin-cent provided more opportunities to be compassionate Thecaveat is that the link between frequency of practice andincrease in self-compassion may not be direct Althoughmindfulness and self-kindness were most often observedonly common humanity improved significantly accordingto our exploratory quantitative analysis Finding commonhumanity in and through a chatbot is also a strong theme inour thematic analysis

Thematic analysis We categorized our data according tothree pillars of self-compassion [32] as displayed in Table 1While all three sub-components were present in both care-receiving and caregiving conditions more instances occurredwith CR Vincent The numbers below a theme (Table 1) arecounts of how many times it occurred in each condition Allquotes below were to CR VincentAs quotes in Table 1 suggest many participants offered

helpful advice to CR Vincent Vincent showed appreciation

Theme QuoteMindfulnessCaregiving 3Care-receiving 25

There are worse things thatcould happen What has hap-pened has happened

Self-kindnessCaregiving 7Care-receiving 21

Go do something fun todaylike watching a movie Staypositive and keep trying untilyou succeed

Common humanityCaregiving 0Care-receiving 11

Everyone makes mistakesJust remember that it can hap-pen to anyone and that itrsquos notyour fault

Table 1 Self-compassion sub-components

with follow-up statements like ldquoyou always make me feel bet-terrdquo Negative counterparts to three pillars of self-compassionwere not strongly present ie self-judgment was detectedfour times for CR Vincent and once for CG Vincent isolationwas noted once for CR Vincent but none for CG Vincentand over-identification was neither present for CG nor CRVincent

For both conditions people were mostly friendly to Vin-cent and there were no swear words or abusive languagedisplayed The most hostile comment was ldquoyoursquove beenpretty dumbrdquo to CR Vincent and we encountered suchldquoput-downsrdquo only twice There were additional topics thatemerged through open thematic analysis They are sum-marized in Table 2 and these themes could also pertain toself-compassion themes (messages to CGVincent are markedwith ldquoCGrdquo and otherwise they were to CR Vincent)

Theme QuotePragmatismCaregiving 0Care-receiving 41

Maybe next time make a bet-ter planning and make sureyoursquove got enough time )

Perspective-takingCaregiving 0Care-receiving 10

I would find a window toclimb in But maybe in yourcase better try to hack into thefolder [] be proud of thebot that you are

Engagement vsDistantiationCaregiving 27 vs 6Care-receiving 5 vs 11

A girl told me she loves meAnd I love her too (CG) vsSorry itrsquos confidential (CG)

Positive vs NegativeCaregiving 74 vs 9Care-receiving 5 vs 2

I was laying in a field full offlowers trying out my newukelele (CG) vs I hate pink

Table 2 Free-input themes

People interacted with CG and CR Vincents differently(Table 2) Giving pragmatic advice to Vincent and taking itsperspective as a chatbot were themes only found in the CRcondition Rather than tending to Vincent by giving emo-tional support participants gave practical advice on what todo better Examples of perspective-taking are recommendingVincent to ldquohack into the folderrdquo or to use ldquobrute forcerdquo tech-niques to gain access participants thought like a chatbot tohelp a chatbotSome participants revealed more personal information

to CG Vincent (theme engagement) and took interest inVincent by asking questions back eg ldquowhat (did you) dofor money before nowrdquo or writing lengthy responses Someshared information was very intimate in nature eg ldquoIrsquomgoing to kiss the girl next to me in 5 secondsrdquo Since CGVincent asked participants to write about themselves thisskewed engagement (the amount of textual response) to-wards CG Vincent Participants distanced themselves fromVincent only a few times by stating that certain informationwas confidential or not showing interest in getting to knowVincent eg ldquosorry I do not know what interests yourdquo Thelast theme on positive vs negative attitude was primarilypresent in the CG condition this theme was mostly aboutattitudes participants had about themselves and their livesnot about Vincent Most participants shared positive lifeevents eg getting an internship cooking something deli-cious Though negative attitudes were minimal they rangedfrom more mundane states eg feeling ldquoawkwardrdquo to moredramatic states eg ldquonot die within 2 weeksrdquoTo summarize Tables 1 and 2 self-compassion sub- com-

ponents were more present with CR Vincent suggestingthat giving compassion to Vincent (or another being) thantowards oneself may be more natural in conversational con-texts And mindfulness most frequently occurred (Table 1)As for emergent themes in Table 2 participants gave prag-matic advice to CR Vincent and often practiced perspective-taking Yet CG Vincent allowed for more self-expression ifparticipants were open to communicate as shown by greaterinstances of engagement and positive remarks about every-day situations In a few instances we detected deeply per-sonal messages on ups and downs of relationships and self-deprecating thoughts Mostly participants shared positivedaily news with CG Vincent and helpful or uplifting remarkswith CR Vincent

Interpretive analysis We now offer a broader interpreta-tion of our data by incorporating participantsrsquo open-endedresponses to an item on the final survey The main theme isbonding between participants and Vincent though not allbonded with Vincent in the same way To explain this weprovide three subthemes that underlie the bonding processwith Vincent Our primary focus was on CR Vincent

Relatability leads to believability Participantsrsquo ability toextend the sense of common humanity to a chatbot touchesupon anthropomorphism CR Vincent was comforted as ifit were a human eg ldquoitrsquos human to make mistakesrdquo (CR)while its problems were addressed to its ldquochatbot worldrdquo egldquocommunicate whatrsquos going on to your fellow chatbotsrdquo (CR)For one participant even Vincentrsquos limitation of having astrict script was anthropomorphized ie ldquoVincent is like thefriend who always speaks about himself and what he haslearned or done and sometimes out of courtesy (not out ofcuriosity) asks how you are doing - but doesnrsquot listen to youranswer or what you actually have to say he just goes onwith his own thingrdquo (CG) Such attributed anthropomorphictraits depended on participantsrsquo willingness take Vincentrsquosperspective as a chatbotCR Vincentrsquos blunders were based on common human

mishaps like being late for a meeting and dealing with un-paid bills (scenarios from [13]) Yet none of our participantsquestioned whether or not a chatbot had meetings to attendor bills to pay Vincentrsquos narrative was on how a chatbotcould be late (new updates took longer than expected) orhow it could have bills (Vincent needs to pay the hostingserver) and our participants went along with imagined sce-narios Vincent faced Rather than questioning the parametersof our scenarios on realism participants thought of how tosolve Vincentrsquos problems within the parameters of a chat-botrsquos world When relevant CR Vincent played up the ironyof having human struggles as a chatbot eg ldquoall I am is lit-erally a piece of code and I failed a programming courserdquoVincent became believable because its struggles were relat-able Granting Vincent human-likness was less literal in howpeople bonded with Vincent Vincent did not try to appearhuman but it socialized with participants about its strugglesthat humans also had People related to Vincentrsquos strugglesand believed that such struggles could arise for chatbotsShared history can lead to attachment Conversations be-

tween people as well as in human-computer interactionbecome shared history over time For instance ldquo[] commu-nicating with Vincent everyday for two weeks builds somekind of habit It makes me notice its presence and absence(which might be good) I think it has a potential to be a goodcompanion and improve the mood especially if someone isfeeling lonelyrdquo (CG) Thus frequent communication with achatbot in a given duration can form expectations ldquoI reallymissed Vincent when he started our conversation laterdquo (CR)The level of attachment for some participants was higherthan others eg after the experiment we saw reactions suchas ldquocan I keep himrdquo (CG)When Vincent prepared participants for its daily good-

byes eg ldquoI have some chatbot things to do Defragmentmy server stack Buy aluminum foil to make fashionablehats withrdquo what was intended as humor can be interpreted

differently ie server defragmentation could be life-or-deathfor a chatbot Some people can be confused worried or evenangered when a chatbot they care about does not respondThus one reaction was ldquothe asshole decided to delete itsstack and when I said itrsquod die it just didnrsquot reply You canrsquotgo making people worried about a freaking chatbotrdquo (CR)People may miss a chatbot that suddenly leaves them orsincerely worry about its well-being This is the positiveand negative aspect of a relatable a chatbot some partic-ipants found common-humanity in Vincent and of thoseparticipants a few possibly related more through empathicdistress rather than through compassion If two weeks canbring about strong signs of attachment longer periods ofinteraction may heighten the level of attachment to differentdegrees and in different ways per person

Emotional reciprocity with chatbots As mentioned beforemost participants were able to respond to CR Vincentrsquos emo-tional displays on a practical level eg recommending howto fix a problem or advising Vincent on how to adjust itsemotions eg telling Vincent to stay positive To add somepeople may not appreciate chatbots demonstrating feelingsOthers may reciprocate or feel comforted by a chatbotrsquos ex-pressed emotions even if a chatbot is perceived as incapableof having emotions The more nuanced point is that Vin-centrsquos display of emotions was noted to bring conflictingfeelings For instance ldquowhen Vincent would show emotions(for example lsquolove talking to yoursquo lsquomiss yoursquo) that would feelweird because I know I am talking to a chatbot and it proba-bly is not that developed that it does have feelings But theusage of such words does feels nice compared to when a hu-man would say them So I had conflicted feelings about thesekind of expressionsrdquo (CG) The participant felt conflictedabout how to process Vincentrsquos emotional outreach

Importantly the participant suggested heshemay bemorecomfortable with a chatbot saying ldquomiss yourdquo than a humanTo conjecture the participant could mean that there was nosocial pressure due to a chatbot not expecting or needinghimher to say ldquoI miss you toordquo People often feel obligatedto respond to sincere emotions of others with similarly va-lenced emotional displays even if they do not feel the samesincere emotions towards them Such pressure may not holdfor technological entities Perhaps to miss someone impliesa certain history in a relationship so to hear that from aperson one met less than two weeks ago may feel awkwardor insincere whereas a chatbot would not be expected toknow or abide by certain social conventions If two peopleknew beforehand they will get to know each other for a maxi-mum duration of two weeks (as our participants knew beforemeeting Vincent) and never be in touch again their emo-tional performance may adjust accordingly The timescalefor intensifying socially acceptable emotional expressions

in human-chatbot interactions and human-human interac-tions may differ The ldquolifespanrdquo of a chatbot is not equatableto a personrsquos lifespan And the distinction between super-ficial vs genuine emotional displays from and to a chatbotis not entirely equatable to emotions people share and re-ciprocate between each other Currently we do not haveestablished norms on how emotions between humans andbots areshould be managed We suggest there may be dis-tinct differences compared to emotions in human-humanrelationships

Discussion and design implicationsCR Vincent adds depth to the CASA paradigm [16 31]mdash notonly do people treat a chatbot as an anthropomorphizedsocial agent but they themselves are affected by a chatbotto the extent that their self-compassion can increase whenthey are compassionate towards a chatbot Brave et alrsquosinsight on embodied conversational agents is that ldquojust aspeople respond to being cared about by other people usersrespond positively to agents that carerdquo [4 p 174] We addthat just as giving care to another human can increase onersquosself-compassion [5] caring for a chatbot can enhance onersquosown self-compassion If the dominant question has beenldquowhat can technology do for usrdquo Vincent demonstrates thatby exploring ldquowhat can we do for technologyrdquo we inad-vertently benefit from technology potentially more so thanwhen we only shape technology to serve us This claim isspecified to bots in the mental health domain and our goalwas to increase self-compassion as a target for well-being[52] rather than to reduce clinically defined symptoms ofpsychological ill-being We present our design implicationsbelow on building chatbots for psychological health carewhich primarily stem from our interpretive analysis Ourimplications are inter-related starting points that should becontextualized for each research and deployment processGive users more closed-inputs or free-input optionsMany

participants felt limited in responses they could give to Vin-cent They wanted to write to Vincent without having anypreset answers or needed more options A recommendationis to use natural language processing for a chatbot whichwill rely less on a pre-planned narrative arc and build moreon what users say This will require a longer developmentperiod The simpler option is to provide users with morefixed responses (three to four) and more opportunities foropen input

Develop a chatbotrsquos story with users Conversational agentscan be powerful storytellers [25 36] even without complexAI To deliver co-storytelling as shared history with interac-tants we suggest designers to create flexible narrative pa-rameters that people can creatively use to relate to a chatbotVincent was able to tell its story but it was less interactive inthat people could follow along with limited reaction options

due to the nature of our experiment There can be additionalcomplexities that designers can add For instance the nar-rative can take a different route depending on which closedinput options users click on We have utilized a limited num-ber of messages called ldquopathsrdquo (supplementary material) thatVincent could give depending on closed input responsesYet this practice did not change Vincentrsquos narrative Givinga chatbot ldquomemoryrdquo be it knowing basic information likenames or a more involved retention of what users say canenhance conversational storytellingTread carefully with emotional expressionsWe suggest a

broader view on what emotions are by considering inter-related emotions that develop over time For example for abot to miss someone assumes a botrsquos happinessenjoymentexperienced during a prior interaction with a user a botrsquosability to feel longing should follow its prior display of joyshared with the user This requires critically formulating in-tentions behind communicative moves [44] of any affectivebot There are several paths for developing emotional dis-plays To list a few (1) offer one type of consistent emotionalexpressions as Vincent did (2) design emotional expressionsthat may be appropriate for different target groups in tan-dem with the implication below and (3) give users controlover how their chatbots ldquofeelrdquo towards them The caveat forthe third recommendation is that the user control over achatbotrsquos emotions may not aid a chatbotrsquos narrative andit also may not be helpful for all users the associated riskis that user-controlled emotions can render a chatbot lessrelatable More specifically the illusion that a chatbot canauthentically care for or be cared by another being requiressome level of perceived independence in how it ldquofeelsrdquo Werecommend designers to engage with the growing field ofaffective computing [41 45] and its discussion on ethics [8]If a chatbotrsquos goal is bettering usersrsquo psychological statesdesigners must ask if an affective bot delivers the intendedtreatment and what ethical boundaries there are in its dis-plays and elicitation of emotions Designers and users couldcontrol a chatbotrsquos emotions but what emotions a chatbotcan elicit in users is not always a priori foreseeable

Tailor chatbots to different target groups Evenwith one con-struct self-compassion we see a variety of ways a chatbotcan be configured To start people with low self-compassionmay benefit the most from Vincent as our exploratory anal-ysis shows This can mean more compassion focused sce-narios rather than neutral scenarios Women are noted toscore lower on self-compassion [51] yet older women experi-ence greater compassion than older men [29] Chatbots thatconsider gender age andor occupation can be potentiallyhelpful for increasing self-compassion To list a few examplesfor reincarnating Vincent a chatbot could be gendered asfemale or non-binary present a proactive version of compas-sion specified for women (see eg Neff [35] on speaking up

and protecting oneself from harm) talk about exam stresswith students or refer to stressful meetings or workplacebullying with employed individuals Rather than assumingthat one-size-fits-all or extreme personalization will workwe suggest designers to first approach targeted groups toclearly understand their needs For instance whether a self-compassion chatbot for all women is as effective or moreeffective than a more targeted chatbot eg at various levelsof intersectionality like race culture age etc should beconsidered given the time and resources that may be avail-able We recommend that based on research uncoveringpossible ways to design a chatbot that suits different needsand wants should be prioritized

Future works and limitationsOur study opened up new questions to be explored An av-enue to investigate is how a chatbotrsquos use of emotional lan-guage influences its interactants We posit that a sufferingchatbot induces less empathic distress than a suffering hu-man and whether or not this is the case needs to be furtherinvestigated especially for chatbots intended to be therapeu-tic helpers An awareness of onersquos own and othersrsquo sufferingwithout overwhelming empathic distress is suggested to bepossible through compassion [18 47] Hence disambiguat-ing compassion from empathic distress is critical in deploy-ing self-compassion chatbots as instantiations of positivecomputing [6] Different configurations of Vincent basedon peoplersquos gender age or occupation could improve theirself-compassion scores more effectively and if and in whatways this holds true warrants further research

There are limitations to consider Our effect size was lowerthan findings from the Woebot study (d = 044) which had34 participants who ldquoself-identified as experiencing symp-toms of depression and anxietyrdquo [15 p 2] and they measuredsymptoms of depression with the PHQ-9 questionnaire notself-compassion Falconer et alrsquos results on self-compassionscores after embodied VR experience also had a higher ef-fect size with the partial eta-squared of 036 (d = 1487) [14]which was based on 15 participants with depression Weworked with a general non-clinical sample and CG Vincentshowed an effect size of d = 007 (N = 34) and CR Vincentrsquoseffect size was d = 02 (N =33) Follow-up studies on self-compassion chatbots can utilize a larger sample and a moregrounded effect size One explanation for the difference ineffect size is that we did not recruit people who were clini-cally or self-proclaimed to be depressed based on the viewthat preventative mental health care can build resilience forpeople in general While Vincent andWoebot [15] share com-monalities the main measurements and targeted populationdiffered And while self-compassion was the measurementfor us and Falconer et al [14] the technology used sample

size and targeted population differed The gain andor main-tenance of healthy self-compassion as pre-emptive care maynot result in a similarly high effect size but can be psycholog-ically beneficial nonetheless More research is necessary tounderstand long-term consequences of a priori preventativecare vs a posteriori treatment of mental health woes

More broadly peoplersquos engagement with Vincent may re-flect both socially desirable reactions such as politeness to-wards machines as social actors [16 31] as well as emotionalempathy ie the ability to ldquofeel forrdquo VincentWe have not yetconcretely looked into other potential contributing factorsin bringing about self-compassion through human-chatbotinteraction Also what is difficult to gauge is the magnitudeof a chatbotrsquos perceived social and emotional complexitybased solely on messaging or text-based conversations Vin-cent lacked embodied communication which means it didnot use non-verbal modalities such as gaze voice or ges-tures that are critical in various social interactions Vincentwas a uni-modal technological entity that can be extendedthrough other complex emotional displays Thus we havenot established how people would engage with other formsof technology like robots with varying degrees and types ofembodiment alongside different combinations of modalitiesUtilizing technology appropriately for mental health carerequires many comparative renditions

5 CONCLUSIONCompassion is a key moral emotion [19] or motivation [6]that deserves to be further explored through positive comput-ing or technology for well-being Self-compassion can helppeoplersquos overall well-being [52] through kindness towardsoneself connectedness to greater humanity andmindfulnessWhile a chatbot is not a panacea for curing psychologicaldifficulties and is not meant to replace professional helpwe demonstrated that caring for a chatbot can help peoplegain greater self-compassion than being cared for by a chat-bot Our quantitative and qualitative analyses suggest thathuman-chatbot interaction is a promising arena for positivecomputing

REFERENCES[1] Anxiety and Depression Association of America (ADAA) 2016 ADAA

Reviewed Mental Health Apps httpsadaaorgmental-health-appsAccessed 2018-06-25

[2] Arthur Aron Elaine N Aron and Danny Smollan 1992 Inclusion ofOther in the Self Scale and the structure of interpersonal closenessJournal of Personality and Social Psychology 63 4 (1992) 596 httpsdoiorg1010370022-3514634596

[3] Virginia Braun and Victoria Clarke 2006 Using thematic analysisin psychology Qualitative Research in Psychology 3 2 (2006) 77ndash101httpsdoiorg1011911478088706qp063oa

[4] Scott Brave Clifford Nass and Kevin Hutchinson 2005 Computersthat care investigating the effects of orientation of emotion exhibited

= 02) between prior (M = 3137 SD = 0613) and post (M= 3257 SD = 0558) scores for self-compassion (normalityassumed for prior (W = 0979 p = 0764) and post (W = 0946p = 0114) scores)We conducted exploratory analyses to better understand

our data Through a repeated measures ANOVA we checkedfor the effect of time prior and post self-compassion scores(F(1 62) = 2768 p = 0101 η2 = 0043) Then we checked forthe effect of condition ie CG or CR (F(1 62) = 075 p = 0785η2 = 0001) and the interaction between time and condition(F(1 62) = 0580 p = 0449 η2 = 0009) None were significantWe additionally checked for time condition timeconditioneffects on the three components of self-compassion self-kindness common humanity and mindfulness Only theeffect of time on common humanity was significant F(1 62)= 6059 p = 0017) The table for all statistical tests is in thesupplementary materialWe further dissected our data by gender because previ-

ous research showed that women may score lower on self-compassion [51] Indeed female participants had lower self-compassion scores (M = 305 SD = 013) than men (M = 326SD = 009) at the start but not significantly so (t (4935) =113 p = 026 d = 029) according an independent unequalvariance t-test (normality assumed) We then compared postand prior scores for men and women Menrsquos self compassionscores increased only by 002 as a difference in means andshowed no significant increase (t (33) = -025 p = 040 d =004) However womenrsquos scores showed a significant differ-ence (t (27)= -206 p = 002) between 305 (SD = 071) as thestarting score and 319 (SD = 065) as a posterior score forself-compassion When we scrutinized the gender differencebetween CG and CR Vincents we noticed a more dramaticdifference Women with CR Vincent showed a highly signifi-cant change (t(13) = -289 p = 0006 d = 077) compared towomen with CG Vincent (t(13) = -033 p = 037 d = 009)and both were normally distributed

We wanted to check if mainly gender was at stake or if itwas simply a difference between low vs high scorers on priorself-compassion levels We thus divided all participants intotwo groups based on the average self-compassion score at thestart 314 Those who scored above this were high scorers(M = 371 SD = 037 N = 28) those who scored below werelow scorers (M = 269 SD = 032 N = 36) We included oneparticipant with the average score in the high-scoring groupand this had no impact on significance reached Low scorersgreatly increased their self-compassion scores in terms ofsignificance (t(35) = -341 p = 00008 d = 057) but highscorers did not show improvements (t(27) = 110 p = 086 d= 018) Yet normality was not assumed for both low-scorers(W = 093 p = 003) and high-scorers (W = 091 p = 002)since we divided a normally distributed group into two Thuswe performed a two-sample Wilcoxon rank-sum test to see

if there was a significant difference between low scorers andhigh scorers which was the case at z = 286 and p = 0004 CRVincent improved low scorersrsquo self-compassion significantly(t(17) = -320 p = 0003 d = 075) compared to a marginalsignificance for CG Vincent (t(17) = -175 p = 005 d = 041)There were normal distributions observed for prior and postself-compassion levels of low and high scorersA potential explanation for why low scorers improved

more than high-scorers is regression to the mean Howeverin published literature the average self-compassion score isbetween 25 and 35 [33] and our low scorers have a prior av-erage self-compassion of 269 If regression to the mean is anexplanation we would also expect high-scorers to end witha lower mean yet this is not the case Our high-scorers hadan average prior score of 371 (above average [33]) and theirscores did not decrease after the experiment This may be aceiling effect The low-scorersrsquo improvement is still there itis a highly significant effect even with the Bonferroni correc-tion for all tests (p = 00008) with the post-hoc power of 097The data supports that Vincent enhanced self-compassionfor low-scorersWe had two additional scales one to check if people per-

ceived CG and CR Vincent in a relatively similar way [4]and the other to see how much participants identified withVincent [2] The survey on participantsrsquo opinion of the agentincluded caring likability trustworthiness intelligence dom-inance and submissiveness as items [4] about Vincent BothCG and CR Vincents were perceived to be fairly analogoussave for dominance (α = 048 CG Vincent M = 2677 SD =0794 CR Vincent M = 2656 SD = 0700) and submissiveness(α = 024 CG Vincent M = 2448 SD = 0559 CR Vincent M= 2396 SD = 0636) though none showed a significant dif-ference between two conditions The the Inclusion of Otherin the Self (IOS) scale [2] displayed that participants moreclosely related to CR Vincent (M = 348 SD = 148) thanCG Vincent (M = 306 SD = 139) but not significantly so (t= -115 p = 025 d = 029) Both conditions were normallydistributed (CG W = 096 p = 034 CR W = 097 p = 064)

To summarize our hypothesis that CG Vincent increasesself-compassion was not supported (p = 0286) but the hy-pothesis that CR Vincent increases self-compassion was sup-ported (p = 0029) Our exploratory analyses captured threeunderlying influences on this finding First our ANOVA testsrevealed that the only significant aspect was time as an inde-pendent variable affecting common humanity one elementof self-compassion (p = 0017) Second gender may be a con-tributing factor with women demonstrating a significantincrease in self-compassion (p = 002) for both conditionscombined but not men (p = 040) To add only CR Vincentdemonstrated a highly significant change for women (p =0006) unlike women who interacted with CG Vincent (p =037) Third regardless of gender those who started out with

a low self-compassion score exhibited the most significantchange (p = 00008) for both conditions together Low-scorersmore significantly improvedwith CR Vincent (p = 0003) thanwith CG Vincent (p = 005)

Put together CR Vincent more effectively increased self-compassion than CG Vincent most likely through a signif-icant change in participantsrsquo sense of common humanitymore so than self-kindness and mindfulness (as our ANOVAtest showed) Finding common humanity can be inclusive ofchatbots Women specifically those with CR Vincent weresignificantly more affected than men However low-scorersof both genders benefited the most compared to high-scorersespecially those with CR Vincent CG and CR Vincents werenot perceived to be significantly different except for a lowersimilarity regarding the dominance-submissive trait Par-ticipants in the CR condition may have felt that Vincent ismore like them (ISO scale) though this difference was notsignificant

Qualitative analysisFor our qualitative analysis we used the corpus of free re-sponses that participants had typed during their interactionswith Vincent We will first present a descriptive analysisof the interactions that people had with CG and CR Vin-cents followed by our thematic analysis [5] and interpretiveanalysis [48] Participantsrsquo responses to CR Vincent were onaverage 9353 words (SD = 4796) and to CG Vincent 11247words (SD = 6399) for two weeks While our data set is notabundant in terms of word count we believe a qualitativelook at how participants interacted with Vincent is valuable

Descriptive analysis CGVincent guided participants throughself-compassion exercises [34] that did not require them toactively voice aspects of self-compassion they mainly hadto read about it This resulted in fewer instances of self-compassion themes in CG Vincent since they only occurredwhen Vincent asked participants to comfort themselves Toadd participantsrsquo willingness to engage with CG Vincentrsquosprobes differed It asked specific questions or provided ashort task for free input eg ldquowrite down a difficulty youhaverdquo Many answers to this were short ldquowake up earlyrdquoor ldquoI often feel alonerdquo Some participants opened up moreldquoI have a difficulty in expressing myself when I am underdifficult situationsrdquo or ldquoI am studying abroad far from homeand family and friends Different culture language edu-cational standardrdquo CG Vincent asked a follow-up questionldquohow did it make you feelrdquo we again got simple answerslike ldquogoodrdquo or ldquonormalrdquo or longer expressions ldquoI feel morerefreshedrdquo or ldquonot really better or worserdquo In other instancesCG Vincent allowed participants to give themselves simpleself-assurance ldquoI can do itrdquo plan for long-term goals ldquoonceI have graduated I can schedule a good routine to target a

fitter and healthier lifestyle Just hang in there a little longerrdquoor dig deeper ldquopeople around me would be happier if I washappierrdquo Thus CG Vincent provided a daily touch point forself-reflection admittance of everyday suffering ldquopep-talkrdquoor deeper self-insight which may or may not directly relateto self-compassion for all participants

In contrast CR Vincent frequently asked for help and con-sequently received many self-compassion related answersThe narrative was the focus for CR Vincent It was able tobecome more vulnerable over time by admitting its own ev-eryday hardships as a chatbot which led it to seek opinion oradvice For example CR Vincent asked ldquowhat do you thinkam I the dumbest bot yoursquove ever seen or what Am I beingtoo hard on myselfrdquo To this participants responded in dif-ferent ways ldquoI think yoursquore the funniest bot that Irsquove everseen mdash yes you are in some situationsrdquo ldquoNo but you shouldexpect (that) a bot is very smartly programmed and know allmdash Maybe I do not knowrdquo or ldquothe world needs bots like youAnd itrsquos usual to get rejected sometimes just keep on goingand yoursquoll find a job soon enoughrdquo However CR Vincentrsquoscries for help did not always result in necessarily compassion-ate replies Many users stuck to pragmatic answers relatedto the topic of the problem Even though all of CR Vincentrsquosscenarios were intended to generate compassion towardsVincent pragmatic replies indicate that not everyone willdemonstrate compassionate responses in every instanceThe difference between CG and CR Vincents is that be-

ing compassionate towards another being in a conversa-tional narrative context is unlike doing guided exerciseson self-compassion about oneself The frequency of con-structing compassionate replies is a way to practice self-compassion users of CR Vincent spent more time practicingself-compassion than those with CG Vincent Therefore CRVincent was more effective than CG Vincent since CR Vin-cent provided more opportunities to be compassionate Thecaveat is that the link between frequency of practice andincrease in self-compassion may not be direct Althoughmindfulness and self-kindness were most often observedonly common humanity improved significantly accordingto our exploratory quantitative analysis Finding commonhumanity in and through a chatbot is also a strong theme inour thematic analysis

Thematic analysis We categorized our data according tothree pillars of self-compassion [32] as displayed in Table 1While all three sub-components were present in both care-receiving and caregiving conditions more instances occurredwith CR Vincent The numbers below a theme (Table 1) arecounts of how many times it occurred in each condition Allquotes below were to CR VincentAs quotes in Table 1 suggest many participants offered

helpful advice to CR Vincent Vincent showed appreciation

Theme QuoteMindfulnessCaregiving 3Care-receiving 25

There are worse things thatcould happen What has hap-pened has happened

Self-kindnessCaregiving 7Care-receiving 21

Go do something fun todaylike watching a movie Staypositive and keep trying untilyou succeed

Common humanityCaregiving 0Care-receiving 11

Everyone makes mistakesJust remember that it can hap-pen to anyone and that itrsquos notyour fault

Table 1 Self-compassion sub-components

with follow-up statements like ldquoyou always make me feel bet-terrdquo Negative counterparts to three pillars of self-compassionwere not strongly present ie self-judgment was detectedfour times for CR Vincent and once for CG Vincent isolationwas noted once for CR Vincent but none for CG Vincentand over-identification was neither present for CG nor CRVincent

For both conditions people were mostly friendly to Vin-cent and there were no swear words or abusive languagedisplayed The most hostile comment was ldquoyoursquove beenpretty dumbrdquo to CR Vincent and we encountered suchldquoput-downsrdquo only twice There were additional topics thatemerged through open thematic analysis They are sum-marized in Table 2 and these themes could also pertain toself-compassion themes (messages to CGVincent are markedwith ldquoCGrdquo and otherwise they were to CR Vincent)

Theme QuotePragmatismCaregiving 0Care-receiving 41

Maybe next time make a bet-ter planning and make sureyoursquove got enough time )

Perspective-takingCaregiving 0Care-receiving 10

I would find a window toclimb in But maybe in yourcase better try to hack into thefolder [] be proud of thebot that you are

Engagement vsDistantiationCaregiving 27 vs 6Care-receiving 5 vs 11

A girl told me she loves meAnd I love her too (CG) vsSorry itrsquos confidential (CG)

Positive vs NegativeCaregiving 74 vs 9Care-receiving 5 vs 2

I was laying in a field full offlowers trying out my newukelele (CG) vs I hate pink

Table 2 Free-input themes

People interacted with CG and CR Vincents differently(Table 2) Giving pragmatic advice to Vincent and taking itsperspective as a chatbot were themes only found in the CRcondition Rather than tending to Vincent by giving emo-tional support participants gave practical advice on what todo better Examples of perspective-taking are recommendingVincent to ldquohack into the folderrdquo or to use ldquobrute forcerdquo tech-niques to gain access participants thought like a chatbot tohelp a chatbotSome participants revealed more personal information

to CG Vincent (theme engagement) and took interest inVincent by asking questions back eg ldquowhat (did you) dofor money before nowrdquo or writing lengthy responses Someshared information was very intimate in nature eg ldquoIrsquomgoing to kiss the girl next to me in 5 secondsrdquo Since CGVincent asked participants to write about themselves thisskewed engagement (the amount of textual response) to-wards CG Vincent Participants distanced themselves fromVincent only a few times by stating that certain informationwas confidential or not showing interest in getting to knowVincent eg ldquosorry I do not know what interests yourdquo Thelast theme on positive vs negative attitude was primarilypresent in the CG condition this theme was mostly aboutattitudes participants had about themselves and their livesnot about Vincent Most participants shared positive lifeevents eg getting an internship cooking something deli-cious Though negative attitudes were minimal they rangedfrom more mundane states eg feeling ldquoawkwardrdquo to moredramatic states eg ldquonot die within 2 weeksrdquoTo summarize Tables 1 and 2 self-compassion sub- com-

ponents were more present with CR Vincent suggestingthat giving compassion to Vincent (or another being) thantowards oneself may be more natural in conversational con-texts And mindfulness most frequently occurred (Table 1)As for emergent themes in Table 2 participants gave prag-matic advice to CR Vincent and often practiced perspective-taking Yet CG Vincent allowed for more self-expression ifparticipants were open to communicate as shown by greaterinstances of engagement and positive remarks about every-day situations In a few instances we detected deeply per-sonal messages on ups and downs of relationships and self-deprecating thoughts Mostly participants shared positivedaily news with CG Vincent and helpful or uplifting remarkswith CR Vincent

Interpretive analysis We now offer a broader interpreta-tion of our data by incorporating participantsrsquo open-endedresponses to an item on the final survey The main theme isbonding between participants and Vincent though not allbonded with Vincent in the same way To explain this weprovide three subthemes that underlie the bonding processwith Vincent Our primary focus was on CR Vincent

Relatability leads to believability Participantsrsquo ability toextend the sense of common humanity to a chatbot touchesupon anthropomorphism CR Vincent was comforted as ifit were a human eg ldquoitrsquos human to make mistakesrdquo (CR)while its problems were addressed to its ldquochatbot worldrdquo egldquocommunicate whatrsquos going on to your fellow chatbotsrdquo (CR)For one participant even Vincentrsquos limitation of having astrict script was anthropomorphized ie ldquoVincent is like thefriend who always speaks about himself and what he haslearned or done and sometimes out of courtesy (not out ofcuriosity) asks how you are doing - but doesnrsquot listen to youranswer or what you actually have to say he just goes onwith his own thingrdquo (CG) Such attributed anthropomorphictraits depended on participantsrsquo willingness take Vincentrsquosperspective as a chatbotCR Vincentrsquos blunders were based on common human

mishaps like being late for a meeting and dealing with un-paid bills (scenarios from [13]) Yet none of our participantsquestioned whether or not a chatbot had meetings to attendor bills to pay Vincentrsquos narrative was on how a chatbotcould be late (new updates took longer than expected) orhow it could have bills (Vincent needs to pay the hostingserver) and our participants went along with imagined sce-narios Vincent faced Rather than questioning the parametersof our scenarios on realism participants thought of how tosolve Vincentrsquos problems within the parameters of a chat-botrsquos world When relevant CR Vincent played up the ironyof having human struggles as a chatbot eg ldquoall I am is lit-erally a piece of code and I failed a programming courserdquoVincent became believable because its struggles were relat-able Granting Vincent human-likness was less literal in howpeople bonded with Vincent Vincent did not try to appearhuman but it socialized with participants about its strugglesthat humans also had People related to Vincentrsquos strugglesand believed that such struggles could arise for chatbotsShared history can lead to attachment Conversations be-

tween people as well as in human-computer interactionbecome shared history over time For instance ldquo[] commu-nicating with Vincent everyday for two weeks builds somekind of habit It makes me notice its presence and absence(which might be good) I think it has a potential to be a goodcompanion and improve the mood especially if someone isfeeling lonelyrdquo (CG) Thus frequent communication with achatbot in a given duration can form expectations ldquoI reallymissed Vincent when he started our conversation laterdquo (CR)The level of attachment for some participants was higherthan others eg after the experiment we saw reactions suchas ldquocan I keep himrdquo (CG)When Vincent prepared participants for its daily good-

byes eg ldquoI have some chatbot things to do Defragmentmy server stack Buy aluminum foil to make fashionablehats withrdquo what was intended as humor can be interpreted

differently ie server defragmentation could be life-or-deathfor a chatbot Some people can be confused worried or evenangered when a chatbot they care about does not respondThus one reaction was ldquothe asshole decided to delete itsstack and when I said itrsquod die it just didnrsquot reply You canrsquotgo making people worried about a freaking chatbotrdquo (CR)People may miss a chatbot that suddenly leaves them orsincerely worry about its well-being This is the positiveand negative aspect of a relatable a chatbot some partic-ipants found common-humanity in Vincent and of thoseparticipants a few possibly related more through empathicdistress rather than through compassion If two weeks canbring about strong signs of attachment longer periods ofinteraction may heighten the level of attachment to differentdegrees and in different ways per person

Emotional reciprocity with chatbots As mentioned beforemost participants were able to respond to CR Vincentrsquos emo-tional displays on a practical level eg recommending howto fix a problem or advising Vincent on how to adjust itsemotions eg telling Vincent to stay positive To add somepeople may not appreciate chatbots demonstrating feelingsOthers may reciprocate or feel comforted by a chatbotrsquos ex-pressed emotions even if a chatbot is perceived as incapableof having emotions The more nuanced point is that Vin-centrsquos display of emotions was noted to bring conflictingfeelings For instance ldquowhen Vincent would show emotions(for example lsquolove talking to yoursquo lsquomiss yoursquo) that would feelweird because I know I am talking to a chatbot and it proba-bly is not that developed that it does have feelings But theusage of such words does feels nice compared to when a hu-man would say them So I had conflicted feelings about thesekind of expressionsrdquo (CG) The participant felt conflictedabout how to process Vincentrsquos emotional outreach

Importantly the participant suggested heshemay bemorecomfortable with a chatbot saying ldquomiss yourdquo than a humanTo conjecture the participant could mean that there was nosocial pressure due to a chatbot not expecting or needinghimher to say ldquoI miss you toordquo People often feel obligatedto respond to sincere emotions of others with similarly va-lenced emotional displays even if they do not feel the samesincere emotions towards them Such pressure may not holdfor technological entities Perhaps to miss someone impliesa certain history in a relationship so to hear that from aperson one met less than two weeks ago may feel awkwardor insincere whereas a chatbot would not be expected toknow or abide by certain social conventions If two peopleknew beforehand they will get to know each other for a maxi-mum duration of two weeks (as our participants knew beforemeeting Vincent) and never be in touch again their emo-tional performance may adjust accordingly The timescalefor intensifying socially acceptable emotional expressions

in human-chatbot interactions and human-human interac-tions may differ The ldquolifespanrdquo of a chatbot is not equatableto a personrsquos lifespan And the distinction between super-ficial vs genuine emotional displays from and to a chatbotis not entirely equatable to emotions people share and re-ciprocate between each other Currently we do not haveestablished norms on how emotions between humans andbots areshould be managed We suggest there may be dis-tinct differences compared to emotions in human-humanrelationships

Discussion and design implicationsCR Vincent adds depth to the CASA paradigm [16 31]mdash notonly do people treat a chatbot as an anthropomorphizedsocial agent but they themselves are affected by a chatbotto the extent that their self-compassion can increase whenthey are compassionate towards a chatbot Brave et alrsquosinsight on embodied conversational agents is that ldquojust aspeople respond to being cared about by other people usersrespond positively to agents that carerdquo [4 p 174] We addthat just as giving care to another human can increase onersquosself-compassion [5] caring for a chatbot can enhance onersquosown self-compassion If the dominant question has beenldquowhat can technology do for usrdquo Vincent demonstrates thatby exploring ldquowhat can we do for technologyrdquo we inad-vertently benefit from technology potentially more so thanwhen we only shape technology to serve us This claim isspecified to bots in the mental health domain and our goalwas to increase self-compassion as a target for well-being[52] rather than to reduce clinically defined symptoms ofpsychological ill-being We present our design implicationsbelow on building chatbots for psychological health carewhich primarily stem from our interpretive analysis Ourimplications are inter-related starting points that should becontextualized for each research and deployment processGive users more closed-inputs or free-input optionsMany

participants felt limited in responses they could give to Vin-cent They wanted to write to Vincent without having anypreset answers or needed more options A recommendationis to use natural language processing for a chatbot whichwill rely less on a pre-planned narrative arc and build moreon what users say This will require a longer developmentperiod The simpler option is to provide users with morefixed responses (three to four) and more opportunities foropen input

Develop a chatbotrsquos story with users Conversational agentscan be powerful storytellers [25 36] even without complexAI To deliver co-storytelling as shared history with interac-tants we suggest designers to create flexible narrative pa-rameters that people can creatively use to relate to a chatbotVincent was able to tell its story but it was less interactive inthat people could follow along with limited reaction options

due to the nature of our experiment There can be additionalcomplexities that designers can add For instance the nar-rative can take a different route depending on which closedinput options users click on We have utilized a limited num-ber of messages called ldquopathsrdquo (supplementary material) thatVincent could give depending on closed input responsesYet this practice did not change Vincentrsquos narrative Givinga chatbot ldquomemoryrdquo be it knowing basic information likenames or a more involved retention of what users say canenhance conversational storytellingTread carefully with emotional expressionsWe suggest a

broader view on what emotions are by considering inter-related emotions that develop over time For example for abot to miss someone assumes a botrsquos happinessenjoymentexperienced during a prior interaction with a user a botrsquosability to feel longing should follow its prior display of joyshared with the user This requires critically formulating in-tentions behind communicative moves [44] of any affectivebot There are several paths for developing emotional dis-plays To list a few (1) offer one type of consistent emotionalexpressions as Vincent did (2) design emotional expressionsthat may be appropriate for different target groups in tan-dem with the implication below and (3) give users controlover how their chatbots ldquofeelrdquo towards them The caveat forthe third recommendation is that the user control over achatbotrsquos emotions may not aid a chatbotrsquos narrative andit also may not be helpful for all users the associated riskis that user-controlled emotions can render a chatbot lessrelatable More specifically the illusion that a chatbot canauthentically care for or be cared by another being requiressome level of perceived independence in how it ldquofeelsrdquo Werecommend designers to engage with the growing field ofaffective computing [41 45] and its discussion on ethics [8]If a chatbotrsquos goal is bettering usersrsquo psychological statesdesigners must ask if an affective bot delivers the intendedtreatment and what ethical boundaries there are in its dis-plays and elicitation of emotions Designers and users couldcontrol a chatbotrsquos emotions but what emotions a chatbotcan elicit in users is not always a priori foreseeable

Tailor chatbots to different target groups Evenwith one con-struct self-compassion we see a variety of ways a chatbotcan be configured To start people with low self-compassionmay benefit the most from Vincent as our exploratory anal-ysis shows This can mean more compassion focused sce-narios rather than neutral scenarios Women are noted toscore lower on self-compassion [51] yet older women experi-ence greater compassion than older men [29] Chatbots thatconsider gender age andor occupation can be potentiallyhelpful for increasing self-compassion To list a few examplesfor reincarnating Vincent a chatbot could be gendered asfemale or non-binary present a proactive version of compas-sion specified for women (see eg Neff [35] on speaking up

and protecting oneself from harm) talk about exam stresswith students or refer to stressful meetings or workplacebullying with employed individuals Rather than assumingthat one-size-fits-all or extreme personalization will workwe suggest designers to first approach targeted groups toclearly understand their needs For instance whether a self-compassion chatbot for all women is as effective or moreeffective than a more targeted chatbot eg at various levelsof intersectionality like race culture age etc should beconsidered given the time and resources that may be avail-able We recommend that based on research uncoveringpossible ways to design a chatbot that suits different needsand wants should be prioritized

Future works and limitationsOur study opened up new questions to be explored An av-enue to investigate is how a chatbotrsquos use of emotional lan-guage influences its interactants We posit that a sufferingchatbot induces less empathic distress than a suffering hu-man and whether or not this is the case needs to be furtherinvestigated especially for chatbots intended to be therapeu-tic helpers An awareness of onersquos own and othersrsquo sufferingwithout overwhelming empathic distress is suggested to bepossible through compassion [18 47] Hence disambiguat-ing compassion from empathic distress is critical in deploy-ing self-compassion chatbots as instantiations of positivecomputing [6] Different configurations of Vincent basedon peoplersquos gender age or occupation could improve theirself-compassion scores more effectively and if and in whatways this holds true warrants further research

There are limitations to consider Our effect size was lowerthan findings from the Woebot study (d = 044) which had34 participants who ldquoself-identified as experiencing symp-toms of depression and anxietyrdquo [15 p 2] and they measuredsymptoms of depression with the PHQ-9 questionnaire notself-compassion Falconer et alrsquos results on self-compassionscores after embodied VR experience also had a higher ef-fect size with the partial eta-squared of 036 (d = 1487) [14]which was based on 15 participants with depression Weworked with a general non-clinical sample and CG Vincentshowed an effect size of d = 007 (N = 34) and CR Vincentrsquoseffect size was d = 02 (N =33) Follow-up studies on self-compassion chatbots can utilize a larger sample and a moregrounded effect size One explanation for the difference ineffect size is that we did not recruit people who were clini-cally or self-proclaimed to be depressed based on the viewthat preventative mental health care can build resilience forpeople in general While Vincent andWoebot [15] share com-monalities the main measurements and targeted populationdiffered And while self-compassion was the measurementfor us and Falconer et al [14] the technology used sample

size and targeted population differed The gain andor main-tenance of healthy self-compassion as pre-emptive care maynot result in a similarly high effect size but can be psycholog-ically beneficial nonetheless More research is necessary tounderstand long-term consequences of a priori preventativecare vs a posteriori treatment of mental health woes

More broadly peoplersquos engagement with Vincent may re-flect both socially desirable reactions such as politeness to-wards machines as social actors [16 31] as well as emotionalempathy ie the ability to ldquofeel forrdquo VincentWe have not yetconcretely looked into other potential contributing factorsin bringing about self-compassion through human-chatbotinteraction Also what is difficult to gauge is the magnitudeof a chatbotrsquos perceived social and emotional complexitybased solely on messaging or text-based conversations Vin-cent lacked embodied communication which means it didnot use non-verbal modalities such as gaze voice or ges-tures that are critical in various social interactions Vincentwas a uni-modal technological entity that can be extendedthrough other complex emotional displays Thus we havenot established how people would engage with other formsof technology like robots with varying degrees and types ofembodiment alongside different combinations of modalitiesUtilizing technology appropriately for mental health carerequires many comparative renditions

5 CONCLUSIONCompassion is a key moral emotion [19] or motivation [6]that deserves to be further explored through positive comput-ing or technology for well-being Self-compassion can helppeoplersquos overall well-being [52] through kindness towardsoneself connectedness to greater humanity andmindfulnessWhile a chatbot is not a panacea for curing psychologicaldifficulties and is not meant to replace professional helpwe demonstrated that caring for a chatbot can help peoplegain greater self-compassion than being cared for by a chat-bot Our quantitative and qualitative analyses suggest thathuman-chatbot interaction is a promising arena for positivecomputing

REFERENCES[1] Anxiety and Depression Association of America (ADAA) 2016 ADAA

Reviewed Mental Health Apps httpsadaaorgmental-health-appsAccessed 2018-06-25

[2] Arthur Aron Elaine N Aron and Danny Smollan 1992 Inclusion ofOther in the Self Scale and the structure of interpersonal closenessJournal of Personality and Social Psychology 63 4 (1992) 596 httpsdoiorg1010370022-3514634596

[3] Virginia Braun and Victoria Clarke 2006 Using thematic analysisin psychology Qualitative Research in Psychology 3 2 (2006) 77ndash101httpsdoiorg1011911478088706qp063oa

[4] Scott Brave Clifford Nass and Kevin Hutchinson 2005 Computersthat care investigating the effects of orientation of emotion exhibited

a low self-compassion score exhibited the most significantchange (p = 00008) for both conditions together Low-scorersmore significantly improvedwith CR Vincent (p = 0003) thanwith CG Vincent (p = 005)

Put together CR Vincent more effectively increased self-compassion than CG Vincent most likely through a signif-icant change in participantsrsquo sense of common humanitymore so than self-kindness and mindfulness (as our ANOVAtest showed) Finding common humanity can be inclusive ofchatbots Women specifically those with CR Vincent weresignificantly more affected than men However low-scorersof both genders benefited the most compared to high-scorersespecially those with CR Vincent CG and CR Vincents werenot perceived to be significantly different except for a lowersimilarity regarding the dominance-submissive trait Par-ticipants in the CR condition may have felt that Vincent ismore like them (ISO scale) though this difference was notsignificant

Qualitative analysisFor our qualitative analysis we used the corpus of free re-sponses that participants had typed during their interactionswith Vincent We will first present a descriptive analysisof the interactions that people had with CG and CR Vin-cents followed by our thematic analysis [5] and interpretiveanalysis [48] Participantsrsquo responses to CR Vincent were onaverage 9353 words (SD = 4796) and to CG Vincent 11247words (SD = 6399) for two weeks While our data set is notabundant in terms of word count we believe a qualitativelook at how participants interacted with Vincent is valuable

Descriptive analysis CGVincent guided participants throughself-compassion exercises [34] that did not require them toactively voice aspects of self-compassion they mainly hadto read about it This resulted in fewer instances of self-compassion themes in CG Vincent since they only occurredwhen Vincent asked participants to comfort themselves Toadd participantsrsquo willingness to engage with CG Vincentrsquosprobes differed It asked specific questions or provided ashort task for free input eg ldquowrite down a difficulty youhaverdquo Many answers to this were short ldquowake up earlyrdquoor ldquoI often feel alonerdquo Some participants opened up moreldquoI have a difficulty in expressing myself when I am underdifficult situationsrdquo or ldquoI am studying abroad far from homeand family and friends Different culture language edu-cational standardrdquo CG Vincent asked a follow-up questionldquohow did it make you feelrdquo we again got simple answerslike ldquogoodrdquo or ldquonormalrdquo or longer expressions ldquoI feel morerefreshedrdquo or ldquonot really better or worserdquo In other instancesCG Vincent allowed participants to give themselves simpleself-assurance ldquoI can do itrdquo plan for long-term goals ldquoonceI have graduated I can schedule a good routine to target a

fitter and healthier lifestyle Just hang in there a little longerrdquoor dig deeper ldquopeople around me would be happier if I washappierrdquo Thus CG Vincent provided a daily touch point forself-reflection admittance of everyday suffering ldquopep-talkrdquoor deeper self-insight which may or may not directly relateto self-compassion for all participants

In contrast CR Vincent frequently asked for help and con-sequently received many self-compassion related answersThe narrative was the focus for CR Vincent It was able tobecome more vulnerable over time by admitting its own ev-eryday hardships as a chatbot which led it to seek opinion oradvice For example CR Vincent asked ldquowhat do you thinkam I the dumbest bot yoursquove ever seen or what Am I beingtoo hard on myselfrdquo To this participants responded in dif-ferent ways ldquoI think yoursquore the funniest bot that Irsquove everseen mdash yes you are in some situationsrdquo ldquoNo but you shouldexpect (that) a bot is very smartly programmed and know allmdash Maybe I do not knowrdquo or ldquothe world needs bots like youAnd itrsquos usual to get rejected sometimes just keep on goingand yoursquoll find a job soon enoughrdquo However CR Vincentrsquoscries for help did not always result in necessarily compassion-ate replies Many users stuck to pragmatic answers relatedto the topic of the problem Even though all of CR Vincentrsquosscenarios were intended to generate compassion towardsVincent pragmatic replies indicate that not everyone willdemonstrate compassionate responses in every instanceThe difference between CG and CR Vincents is that be-

ing compassionate towards another being in a conversa-tional narrative context is unlike doing guided exerciseson self-compassion about oneself The frequency of con-structing compassionate replies is a way to practice self-compassion users of CR Vincent spent more time practicingself-compassion than those with CG Vincent Therefore CRVincent was more effective than CG Vincent since CR Vin-cent provided more opportunities to be compassionate Thecaveat is that the link between frequency of practice andincrease in self-compassion may not be direct Althoughmindfulness and self-kindness were most often observedonly common humanity improved significantly accordingto our exploratory quantitative analysis Finding commonhumanity in and through a chatbot is also a strong theme inour thematic analysis

Thematic analysis We categorized our data according tothree pillars of self-compassion [32] as displayed in Table 1While all three sub-components were present in both care-receiving and caregiving conditions more instances occurredwith CR Vincent The numbers below a theme (Table 1) arecounts of how many times it occurred in each condition Allquotes below were to CR VincentAs quotes in Table 1 suggest many participants offered

helpful advice to CR Vincent Vincent showed appreciation

Theme QuoteMindfulnessCaregiving 3Care-receiving 25

There are worse things thatcould happen What has hap-pened has happened

Self-kindnessCaregiving 7Care-receiving 21

Go do something fun todaylike watching a movie Staypositive and keep trying untilyou succeed

Common humanityCaregiving 0Care-receiving 11

Everyone makes mistakesJust remember that it can hap-pen to anyone and that itrsquos notyour fault

Table 1 Self-compassion sub-components

with follow-up statements like ldquoyou always make me feel bet-terrdquo Negative counterparts to three pillars of self-compassionwere not strongly present ie self-judgment was detectedfour times for CR Vincent and once for CG Vincent isolationwas noted once for CR Vincent but none for CG Vincentand over-identification was neither present for CG nor CRVincent

For both conditions people were mostly friendly to Vin-cent and there were no swear words or abusive languagedisplayed The most hostile comment was ldquoyoursquove beenpretty dumbrdquo to CR Vincent and we encountered suchldquoput-downsrdquo only twice There were additional topics thatemerged through open thematic analysis They are sum-marized in Table 2 and these themes could also pertain toself-compassion themes (messages to CGVincent are markedwith ldquoCGrdquo and otherwise they were to CR Vincent)

Theme QuotePragmatismCaregiving 0Care-receiving 41

Maybe next time make a bet-ter planning and make sureyoursquove got enough time )

Perspective-takingCaregiving 0Care-receiving 10

I would find a window toclimb in But maybe in yourcase better try to hack into thefolder [] be proud of thebot that you are

Engagement vsDistantiationCaregiving 27 vs 6Care-receiving 5 vs 11

A girl told me she loves meAnd I love her too (CG) vsSorry itrsquos confidential (CG)

Positive vs NegativeCaregiving 74 vs 9Care-receiving 5 vs 2

I was laying in a field full offlowers trying out my newukelele (CG) vs I hate pink

Table 2 Free-input themes

People interacted with CG and CR Vincents differently(Table 2) Giving pragmatic advice to Vincent and taking itsperspective as a chatbot were themes only found in the CRcondition Rather than tending to Vincent by giving emo-tional support participants gave practical advice on what todo better Examples of perspective-taking are recommendingVincent to ldquohack into the folderrdquo or to use ldquobrute forcerdquo tech-niques to gain access participants thought like a chatbot tohelp a chatbotSome participants revealed more personal information

to CG Vincent (theme engagement) and took interest inVincent by asking questions back eg ldquowhat (did you) dofor money before nowrdquo or writing lengthy responses Someshared information was very intimate in nature eg ldquoIrsquomgoing to kiss the girl next to me in 5 secondsrdquo Since CGVincent asked participants to write about themselves thisskewed engagement (the amount of textual response) to-wards CG Vincent Participants distanced themselves fromVincent only a few times by stating that certain informationwas confidential or not showing interest in getting to knowVincent eg ldquosorry I do not know what interests yourdquo Thelast theme on positive vs negative attitude was primarilypresent in the CG condition this theme was mostly aboutattitudes participants had about themselves and their livesnot about Vincent Most participants shared positive lifeevents eg getting an internship cooking something deli-cious Though negative attitudes were minimal they rangedfrom more mundane states eg feeling ldquoawkwardrdquo to moredramatic states eg ldquonot die within 2 weeksrdquoTo summarize Tables 1 and 2 self-compassion sub- com-

ponents were more present with CR Vincent suggestingthat giving compassion to Vincent (or another being) thantowards oneself may be more natural in conversational con-texts And mindfulness most frequently occurred (Table 1)As for emergent themes in Table 2 participants gave prag-matic advice to CR Vincent and often practiced perspective-taking Yet CG Vincent allowed for more self-expression ifparticipants were open to communicate as shown by greaterinstances of engagement and positive remarks about every-day situations In a few instances we detected deeply per-sonal messages on ups and downs of relationships and self-deprecating thoughts Mostly participants shared positivedaily news with CG Vincent and helpful or uplifting remarkswith CR Vincent

Interpretive analysis We now offer a broader interpreta-tion of our data by incorporating participantsrsquo open-endedresponses to an item on the final survey The main theme isbonding between participants and Vincent though not allbonded with Vincent in the same way To explain this weprovide three subthemes that underlie the bonding processwith Vincent Our primary focus was on CR Vincent

Relatability leads to believability Participantsrsquo ability toextend the sense of common humanity to a chatbot touchesupon anthropomorphism CR Vincent was comforted as ifit were a human eg ldquoitrsquos human to make mistakesrdquo (CR)while its problems were addressed to its ldquochatbot worldrdquo egldquocommunicate whatrsquos going on to your fellow chatbotsrdquo (CR)For one participant even Vincentrsquos limitation of having astrict script was anthropomorphized ie ldquoVincent is like thefriend who always speaks about himself and what he haslearned or done and sometimes out of courtesy (not out ofcuriosity) asks how you are doing - but doesnrsquot listen to youranswer or what you actually have to say he just goes onwith his own thingrdquo (CG) Such attributed anthropomorphictraits depended on participantsrsquo willingness take Vincentrsquosperspective as a chatbotCR Vincentrsquos blunders were based on common human

mishaps like being late for a meeting and dealing with un-paid bills (scenarios from [13]) Yet none of our participantsquestioned whether or not a chatbot had meetings to attendor bills to pay Vincentrsquos narrative was on how a chatbotcould be late (new updates took longer than expected) orhow it could have bills (Vincent needs to pay the hostingserver) and our participants went along with imagined sce-narios Vincent faced Rather than questioning the parametersof our scenarios on realism participants thought of how tosolve Vincentrsquos problems within the parameters of a chat-botrsquos world When relevant CR Vincent played up the ironyof having human struggles as a chatbot eg ldquoall I am is lit-erally a piece of code and I failed a programming courserdquoVincent became believable because its struggles were relat-able Granting Vincent human-likness was less literal in howpeople bonded with Vincent Vincent did not try to appearhuman but it socialized with participants about its strugglesthat humans also had People related to Vincentrsquos strugglesand believed that such struggles could arise for chatbotsShared history can lead to attachment Conversations be-

tween people as well as in human-computer interactionbecome shared history over time For instance ldquo[] commu-nicating with Vincent everyday for two weeks builds somekind of habit It makes me notice its presence and absence(which might be good) I think it has a potential to be a goodcompanion and improve the mood especially if someone isfeeling lonelyrdquo (CG) Thus frequent communication with achatbot in a given duration can form expectations ldquoI reallymissed Vincent when he started our conversation laterdquo (CR)The level of attachment for some participants was higherthan others eg after the experiment we saw reactions suchas ldquocan I keep himrdquo (CG)When Vincent prepared participants for its daily good-

byes eg ldquoI have some chatbot things to do Defragmentmy server stack Buy aluminum foil to make fashionablehats withrdquo what was intended as humor can be interpreted

differently ie server defragmentation could be life-or-deathfor a chatbot Some people can be confused worried or evenangered when a chatbot they care about does not respondThus one reaction was ldquothe asshole decided to delete itsstack and when I said itrsquod die it just didnrsquot reply You canrsquotgo making people worried about a freaking chatbotrdquo (CR)People may miss a chatbot that suddenly leaves them orsincerely worry about its well-being This is the positiveand negative aspect of a relatable a chatbot some partic-ipants found common-humanity in Vincent and of thoseparticipants a few possibly related more through empathicdistress rather than through compassion If two weeks canbring about strong signs of attachment longer periods ofinteraction may heighten the level of attachment to differentdegrees and in different ways per person

Emotional reciprocity with chatbots As mentioned beforemost participants were able to respond to CR Vincentrsquos emo-tional displays on a practical level eg recommending howto fix a problem or advising Vincent on how to adjust itsemotions eg telling Vincent to stay positive To add somepeople may not appreciate chatbots demonstrating feelingsOthers may reciprocate or feel comforted by a chatbotrsquos ex-pressed emotions even if a chatbot is perceived as incapableof having emotions The more nuanced point is that Vin-centrsquos display of emotions was noted to bring conflictingfeelings For instance ldquowhen Vincent would show emotions(for example lsquolove talking to yoursquo lsquomiss yoursquo) that would feelweird because I know I am talking to a chatbot and it proba-bly is not that developed that it does have feelings But theusage of such words does feels nice compared to when a hu-man would say them So I had conflicted feelings about thesekind of expressionsrdquo (CG) The participant felt conflictedabout how to process Vincentrsquos emotional outreach

Importantly the participant suggested heshemay bemorecomfortable with a chatbot saying ldquomiss yourdquo than a humanTo conjecture the participant could mean that there was nosocial pressure due to a chatbot not expecting or needinghimher to say ldquoI miss you toordquo People often feel obligatedto respond to sincere emotions of others with similarly va-lenced emotional displays even if they do not feel the samesincere emotions towards them Such pressure may not holdfor technological entities Perhaps to miss someone impliesa certain history in a relationship so to hear that from aperson one met less than two weeks ago may feel awkwardor insincere whereas a chatbot would not be expected toknow or abide by certain social conventions If two peopleknew beforehand they will get to know each other for a maxi-mum duration of two weeks (as our participants knew beforemeeting Vincent) and never be in touch again their emo-tional performance may adjust accordingly The timescalefor intensifying socially acceptable emotional expressions

in human-chatbot interactions and human-human interac-tions may differ The ldquolifespanrdquo of a chatbot is not equatableto a personrsquos lifespan And the distinction between super-ficial vs genuine emotional displays from and to a chatbotis not entirely equatable to emotions people share and re-ciprocate between each other Currently we do not haveestablished norms on how emotions between humans andbots areshould be managed We suggest there may be dis-tinct differences compared to emotions in human-humanrelationships

Discussion and design implicationsCR Vincent adds depth to the CASA paradigm [16 31]mdash notonly do people treat a chatbot as an anthropomorphizedsocial agent but they themselves are affected by a chatbotto the extent that their self-compassion can increase whenthey are compassionate towards a chatbot Brave et alrsquosinsight on embodied conversational agents is that ldquojust aspeople respond to being cared about by other people usersrespond positively to agents that carerdquo [4 p 174] We addthat just as giving care to another human can increase onersquosself-compassion [5] caring for a chatbot can enhance onersquosown self-compassion If the dominant question has beenldquowhat can technology do for usrdquo Vincent demonstrates thatby exploring ldquowhat can we do for technologyrdquo we inad-vertently benefit from technology potentially more so thanwhen we only shape technology to serve us This claim isspecified to bots in the mental health domain and our goalwas to increase self-compassion as a target for well-being[52] rather than to reduce clinically defined symptoms ofpsychological ill-being We present our design implicationsbelow on building chatbots for psychological health carewhich primarily stem from our interpretive analysis Ourimplications are inter-related starting points that should becontextualized for each research and deployment processGive users more closed-inputs or free-input optionsMany

participants felt limited in responses they could give to Vin-cent They wanted to write to Vincent without having anypreset answers or needed more options A recommendationis to use natural language processing for a chatbot whichwill rely less on a pre-planned narrative arc and build moreon what users say This will require a longer developmentperiod The simpler option is to provide users with morefixed responses (three to four) and more opportunities foropen input

Develop a chatbotrsquos story with users Conversational agentscan be powerful storytellers [25 36] even without complexAI To deliver co-storytelling as shared history with interac-tants we suggest designers to create flexible narrative pa-rameters that people can creatively use to relate to a chatbotVincent was able to tell its story but it was less interactive inthat people could follow along with limited reaction options

due to the nature of our experiment There can be additionalcomplexities that designers can add For instance the nar-rative can take a different route depending on which closedinput options users click on We have utilized a limited num-ber of messages called ldquopathsrdquo (supplementary material) thatVincent could give depending on closed input responsesYet this practice did not change Vincentrsquos narrative Givinga chatbot ldquomemoryrdquo be it knowing basic information likenames or a more involved retention of what users say canenhance conversational storytellingTread carefully with emotional expressionsWe suggest a

broader view on what emotions are by considering inter-related emotions that develop over time For example for abot to miss someone assumes a botrsquos happinessenjoymentexperienced during a prior interaction with a user a botrsquosability to feel longing should follow its prior display of joyshared with the user This requires critically formulating in-tentions behind communicative moves [44] of any affectivebot There are several paths for developing emotional dis-plays To list a few (1) offer one type of consistent emotionalexpressions as Vincent did (2) design emotional expressionsthat may be appropriate for different target groups in tan-dem with the implication below and (3) give users controlover how their chatbots ldquofeelrdquo towards them The caveat forthe third recommendation is that the user control over achatbotrsquos emotions may not aid a chatbotrsquos narrative andit also may not be helpful for all users the associated riskis that user-controlled emotions can render a chatbot lessrelatable More specifically the illusion that a chatbot canauthentically care for or be cared by another being requiressome level of perceived independence in how it ldquofeelsrdquo Werecommend designers to engage with the growing field ofaffective computing [41 45] and its discussion on ethics [8]If a chatbotrsquos goal is bettering usersrsquo psychological statesdesigners must ask if an affective bot delivers the intendedtreatment and what ethical boundaries there are in its dis-plays and elicitation of emotions Designers and users couldcontrol a chatbotrsquos emotions but what emotions a chatbotcan elicit in users is not always a priori foreseeable

Tailor chatbots to different target groups Evenwith one con-struct self-compassion we see a variety of ways a chatbotcan be configured To start people with low self-compassionmay benefit the most from Vincent as our exploratory anal-ysis shows This can mean more compassion focused sce-narios rather than neutral scenarios Women are noted toscore lower on self-compassion [51] yet older women experi-ence greater compassion than older men [29] Chatbots thatconsider gender age andor occupation can be potentiallyhelpful for increasing self-compassion To list a few examplesfor reincarnating Vincent a chatbot could be gendered asfemale or non-binary present a proactive version of compas-sion specified for women (see eg Neff [35] on speaking up

and protecting oneself from harm) talk about exam stresswith students or refer to stressful meetings or workplacebullying with employed individuals Rather than assumingthat one-size-fits-all or extreme personalization will workwe suggest designers to first approach targeted groups toclearly understand their needs For instance whether a self-compassion chatbot for all women is as effective or moreeffective than a more targeted chatbot eg at various levelsof intersectionality like race culture age etc should beconsidered given the time and resources that may be avail-able We recommend that based on research uncoveringpossible ways to design a chatbot that suits different needsand wants should be prioritized

Future works and limitationsOur study opened up new questions to be explored An av-enue to investigate is how a chatbotrsquos use of emotional lan-guage influences its interactants We posit that a sufferingchatbot induces less empathic distress than a suffering hu-man and whether or not this is the case needs to be furtherinvestigated especially for chatbots intended to be therapeu-tic helpers An awareness of onersquos own and othersrsquo sufferingwithout overwhelming empathic distress is suggested to bepossible through compassion [18 47] Hence disambiguat-ing compassion from empathic distress is critical in deploy-ing self-compassion chatbots as instantiations of positivecomputing [6] Different configurations of Vincent basedon peoplersquos gender age or occupation could improve theirself-compassion scores more effectively and if and in whatways this holds true warrants further research

There are limitations to consider Our effect size was lowerthan findings from the Woebot study (d = 044) which had34 participants who ldquoself-identified as experiencing symp-toms of depression and anxietyrdquo [15 p 2] and they measuredsymptoms of depression with the PHQ-9 questionnaire notself-compassion Falconer et alrsquos results on self-compassionscores after embodied VR experience also had a higher ef-fect size with the partial eta-squared of 036 (d = 1487) [14]which was based on 15 participants with depression Weworked with a general non-clinical sample and CG Vincentshowed an effect size of d = 007 (N = 34) and CR Vincentrsquoseffect size was d = 02 (N =33) Follow-up studies on self-compassion chatbots can utilize a larger sample and a moregrounded effect size One explanation for the difference ineffect size is that we did not recruit people who were clini-cally or self-proclaimed to be depressed based on the viewthat preventative mental health care can build resilience forpeople in general While Vincent andWoebot [15] share com-monalities the main measurements and targeted populationdiffered And while self-compassion was the measurementfor us and Falconer et al [14] the technology used sample

size and targeted population differed The gain andor main-tenance of healthy self-compassion as pre-emptive care maynot result in a similarly high effect size but can be psycholog-ically beneficial nonetheless More research is necessary tounderstand long-term consequences of a priori preventativecare vs a posteriori treatment of mental health woes

More broadly peoplersquos engagement with Vincent may re-flect both socially desirable reactions such as politeness to-wards machines as social actors [16 31] as well as emotionalempathy ie the ability to ldquofeel forrdquo VincentWe have not yetconcretely looked into other potential contributing factorsin bringing about self-compassion through human-chatbotinteraction Also what is difficult to gauge is the magnitudeof a chatbotrsquos perceived social and emotional complexitybased solely on messaging or text-based conversations Vin-cent lacked embodied communication which means it didnot use non-verbal modalities such as gaze voice or ges-tures that are critical in various social interactions Vincentwas a uni-modal technological entity that can be extendedthrough other complex emotional displays Thus we havenot established how people would engage with other formsof technology like robots with varying degrees and types ofembodiment alongside different combinations of modalitiesUtilizing technology appropriately for mental health carerequires many comparative renditions

5 CONCLUSIONCompassion is a key moral emotion [19] or motivation [6]that deserves to be further explored through positive comput-ing or technology for well-being Self-compassion can helppeoplersquos overall well-being [52] through kindness towardsoneself connectedness to greater humanity andmindfulnessWhile a chatbot is not a panacea for curing psychologicaldifficulties and is not meant to replace professional helpwe demonstrated that caring for a chatbot can help peoplegain greater self-compassion than being cared for by a chat-bot Our quantitative and qualitative analyses suggest thathuman-chatbot interaction is a promising arena for positivecomputing

REFERENCES[1] Anxiety and Depression Association of America (ADAA) 2016 ADAA

Reviewed Mental Health Apps httpsadaaorgmental-health-appsAccessed 2018-06-25

[2] Arthur Aron Elaine N Aron and Danny Smollan 1992 Inclusion ofOther in the Self Scale and the structure of interpersonal closenessJournal of Personality and Social Psychology 63 4 (1992) 596 httpsdoiorg1010370022-3514634596

[3] Virginia Braun and Victoria Clarke 2006 Using thematic analysisin psychology Qualitative Research in Psychology 3 2 (2006) 77ndash101httpsdoiorg1011911478088706qp063oa

[4] Scott Brave Clifford Nass and Kevin Hutchinson 2005 Computersthat care investigating the effects of orientation of emotion exhibited

Theme QuoteMindfulnessCaregiving 3Care-receiving 25

There are worse things thatcould happen What has hap-pened has happened

Self-kindnessCaregiving 7Care-receiving 21

Go do something fun todaylike watching a movie Staypositive and keep trying untilyou succeed

Common humanityCaregiving 0Care-receiving 11

Everyone makes mistakesJust remember that it can hap-pen to anyone and that itrsquos notyour fault

Table 1 Self-compassion sub-components

with follow-up statements like ldquoyou always make me feel bet-terrdquo Negative counterparts to three pillars of self-compassionwere not strongly present ie self-judgment was detectedfour times for CR Vincent and once for CG Vincent isolationwas noted once for CR Vincent but none for CG Vincentand over-identification was neither present for CG nor CRVincent

For both conditions people were mostly friendly to Vin-cent and there were no swear words or abusive languagedisplayed The most hostile comment was ldquoyoursquove beenpretty dumbrdquo to CR Vincent and we encountered suchldquoput-downsrdquo only twice There were additional topics thatemerged through open thematic analysis They are sum-marized in Table 2 and these themes could also pertain toself-compassion themes (messages to CGVincent are markedwith ldquoCGrdquo and otherwise they were to CR Vincent)

Theme QuotePragmatismCaregiving 0Care-receiving 41

Maybe next time make a bet-ter planning and make sureyoursquove got enough time )

Perspective-takingCaregiving 0Care-receiving 10

I would find a window toclimb in But maybe in yourcase better try to hack into thefolder [] be proud of thebot that you are

Engagement vsDistantiationCaregiving 27 vs 6Care-receiving 5 vs 11

A girl told me she loves meAnd I love her too (CG) vsSorry itrsquos confidential (CG)

Positive vs NegativeCaregiving 74 vs 9Care-receiving 5 vs 2

I was laying in a field full offlowers trying out my newukelele (CG) vs I hate pink

Table 2 Free-input themes

People interacted with CG and CR Vincents differently(Table 2) Giving pragmatic advice to Vincent and taking itsperspective as a chatbot were themes only found in the CRcondition Rather than tending to Vincent by giving emo-tional support participants gave practical advice on what todo better Examples of perspective-taking are recommendingVincent to ldquohack into the folderrdquo or to use ldquobrute forcerdquo tech-niques to gain access participants thought like a chatbot tohelp a chatbotSome participants revealed more personal information

to CG Vincent (theme engagement) and took interest inVincent by asking questions back eg ldquowhat (did you) dofor money before nowrdquo or writing lengthy responses Someshared information was very intimate in nature eg ldquoIrsquomgoing to kiss the girl next to me in 5 secondsrdquo Since CGVincent asked participants to write about themselves thisskewed engagement (the amount of textual response) to-wards CG Vincent Participants distanced themselves fromVincent only a few times by stating that certain informationwas confidential or not showing interest in getting to knowVincent eg ldquosorry I do not know what interests yourdquo Thelast theme on positive vs negative attitude was primarilypresent in the CG condition this theme was mostly aboutattitudes participants had about themselves and their livesnot about Vincent Most participants shared positive lifeevents eg getting an internship cooking something deli-cious Though negative attitudes were minimal they rangedfrom more mundane states eg feeling ldquoawkwardrdquo to moredramatic states eg ldquonot die within 2 weeksrdquoTo summarize Tables 1 and 2 self-compassion sub- com-

ponents were more present with CR Vincent suggestingthat giving compassion to Vincent (or another being) thantowards oneself may be more natural in conversational con-texts And mindfulness most frequently occurred (Table 1)As for emergent themes in Table 2 participants gave prag-matic advice to CR Vincent and often practiced perspective-taking Yet CG Vincent allowed for more self-expression ifparticipants were open to communicate as shown by greaterinstances of engagement and positive remarks about every-day situations In a few instances we detected deeply per-sonal messages on ups and downs of relationships and self-deprecating thoughts Mostly participants shared positivedaily news with CG Vincent and helpful or uplifting remarkswith CR Vincent

Interpretive analysis We now offer a broader interpreta-tion of our data by incorporating participantsrsquo open-endedresponses to an item on the final survey The main theme isbonding between participants and Vincent though not allbonded with Vincent in the same way To explain this weprovide three subthemes that underlie the bonding processwith Vincent Our primary focus was on CR Vincent

Relatability leads to believability Participantsrsquo ability toextend the sense of common humanity to a chatbot touchesupon anthropomorphism CR Vincent was comforted as ifit were a human eg ldquoitrsquos human to make mistakesrdquo (CR)while its problems were addressed to its ldquochatbot worldrdquo egldquocommunicate whatrsquos going on to your fellow chatbotsrdquo (CR)For one participant even Vincentrsquos limitation of having astrict script was anthropomorphized ie ldquoVincent is like thefriend who always speaks about himself and what he haslearned or done and sometimes out of courtesy (not out ofcuriosity) asks how you are doing - but doesnrsquot listen to youranswer or what you actually have to say he just goes onwith his own thingrdquo (CG) Such attributed anthropomorphictraits depended on participantsrsquo willingness take Vincentrsquosperspective as a chatbotCR Vincentrsquos blunders were based on common human

mishaps like being late for a meeting and dealing with un-paid bills (scenarios from [13]) Yet none of our participantsquestioned whether or not a chatbot had meetings to attendor bills to pay Vincentrsquos narrative was on how a chatbotcould be late (new updates took longer than expected) orhow it could have bills (Vincent needs to pay the hostingserver) and our participants went along with imagined sce-narios Vincent faced Rather than questioning the parametersof our scenarios on realism participants thought of how tosolve Vincentrsquos problems within the parameters of a chat-botrsquos world When relevant CR Vincent played up the ironyof having human struggles as a chatbot eg ldquoall I am is lit-erally a piece of code and I failed a programming courserdquoVincent became believable because its struggles were relat-able Granting Vincent human-likness was less literal in howpeople bonded with Vincent Vincent did not try to appearhuman but it socialized with participants about its strugglesthat humans also had People related to Vincentrsquos strugglesand believed that such struggles could arise for chatbotsShared history can lead to attachment Conversations be-

tween people as well as in human-computer interactionbecome shared history over time For instance ldquo[] commu-nicating with Vincent everyday for two weeks builds somekind of habit It makes me notice its presence and absence(which might be good) I think it has a potential to be a goodcompanion and improve the mood especially if someone isfeeling lonelyrdquo (CG) Thus frequent communication with achatbot in a given duration can form expectations ldquoI reallymissed Vincent when he started our conversation laterdquo (CR)The level of attachment for some participants was higherthan others eg after the experiment we saw reactions suchas ldquocan I keep himrdquo (CG)When Vincent prepared participants for its daily good-

byes eg ldquoI have some chatbot things to do Defragmentmy server stack Buy aluminum foil to make fashionablehats withrdquo what was intended as humor can be interpreted

differently ie server defragmentation could be life-or-deathfor a chatbot Some people can be confused worried or evenangered when a chatbot they care about does not respondThus one reaction was ldquothe asshole decided to delete itsstack and when I said itrsquod die it just didnrsquot reply You canrsquotgo making people worried about a freaking chatbotrdquo (CR)People may miss a chatbot that suddenly leaves them orsincerely worry about its well-being This is the positiveand negative aspect of a relatable a chatbot some partic-ipants found common-humanity in Vincent and of thoseparticipants a few possibly related more through empathicdistress rather than through compassion If two weeks canbring about strong signs of attachment longer periods ofinteraction may heighten the level of attachment to differentdegrees and in different ways per person

Emotional reciprocity with chatbots As mentioned beforemost participants were able to respond to CR Vincentrsquos emo-tional displays on a practical level eg recommending howto fix a problem or advising Vincent on how to adjust itsemotions eg telling Vincent to stay positive To add somepeople may not appreciate chatbots demonstrating feelingsOthers may reciprocate or feel comforted by a chatbotrsquos ex-pressed emotions even if a chatbot is perceived as incapableof having emotions The more nuanced point is that Vin-centrsquos display of emotions was noted to bring conflictingfeelings For instance ldquowhen Vincent would show emotions(for example lsquolove talking to yoursquo lsquomiss yoursquo) that would feelweird because I know I am talking to a chatbot and it proba-bly is not that developed that it does have feelings But theusage of such words does feels nice compared to when a hu-man would say them So I had conflicted feelings about thesekind of expressionsrdquo (CG) The participant felt conflictedabout how to process Vincentrsquos emotional outreach

Importantly the participant suggested heshemay bemorecomfortable with a chatbot saying ldquomiss yourdquo than a humanTo conjecture the participant could mean that there was nosocial pressure due to a chatbot not expecting or needinghimher to say ldquoI miss you toordquo People often feel obligatedto respond to sincere emotions of others with similarly va-lenced emotional displays even if they do not feel the samesincere emotions towards them Such pressure may not holdfor technological entities Perhaps to miss someone impliesa certain history in a relationship so to hear that from aperson one met less than two weeks ago may feel awkwardor insincere whereas a chatbot would not be expected toknow or abide by certain social conventions If two peopleknew beforehand they will get to know each other for a maxi-mum duration of two weeks (as our participants knew beforemeeting Vincent) and never be in touch again their emo-tional performance may adjust accordingly The timescalefor intensifying socially acceptable emotional expressions

in human-chatbot interactions and human-human interac-tions may differ The ldquolifespanrdquo of a chatbot is not equatableto a personrsquos lifespan And the distinction between super-ficial vs genuine emotional displays from and to a chatbotis not entirely equatable to emotions people share and re-ciprocate between each other Currently we do not haveestablished norms on how emotions between humans andbots areshould be managed We suggest there may be dis-tinct differences compared to emotions in human-humanrelationships

Discussion and design implicationsCR Vincent adds depth to the CASA paradigm [16 31]mdash notonly do people treat a chatbot as an anthropomorphizedsocial agent but they themselves are affected by a chatbotto the extent that their self-compassion can increase whenthey are compassionate towards a chatbot Brave et alrsquosinsight on embodied conversational agents is that ldquojust aspeople respond to being cared about by other people usersrespond positively to agents that carerdquo [4 p 174] We addthat just as giving care to another human can increase onersquosself-compassion [5] caring for a chatbot can enhance onersquosown self-compassion If the dominant question has beenldquowhat can technology do for usrdquo Vincent demonstrates thatby exploring ldquowhat can we do for technologyrdquo we inad-vertently benefit from technology potentially more so thanwhen we only shape technology to serve us This claim isspecified to bots in the mental health domain and our goalwas to increase self-compassion as a target for well-being[52] rather than to reduce clinically defined symptoms ofpsychological ill-being We present our design implicationsbelow on building chatbots for psychological health carewhich primarily stem from our interpretive analysis Ourimplications are inter-related starting points that should becontextualized for each research and deployment processGive users more closed-inputs or free-input optionsMany

participants felt limited in responses they could give to Vin-cent They wanted to write to Vincent without having anypreset answers or needed more options A recommendationis to use natural language processing for a chatbot whichwill rely less on a pre-planned narrative arc and build moreon what users say This will require a longer developmentperiod The simpler option is to provide users with morefixed responses (three to four) and more opportunities foropen input

Develop a chatbotrsquos story with users Conversational agentscan be powerful storytellers [25 36] even without complexAI To deliver co-storytelling as shared history with interac-tants we suggest designers to create flexible narrative pa-rameters that people can creatively use to relate to a chatbotVincent was able to tell its story but it was less interactive inthat people could follow along with limited reaction options

due to the nature of our experiment There can be additionalcomplexities that designers can add For instance the nar-rative can take a different route depending on which closedinput options users click on We have utilized a limited num-ber of messages called ldquopathsrdquo (supplementary material) thatVincent could give depending on closed input responsesYet this practice did not change Vincentrsquos narrative Givinga chatbot ldquomemoryrdquo be it knowing basic information likenames or a more involved retention of what users say canenhance conversational storytellingTread carefully with emotional expressionsWe suggest a

broader view on what emotions are by considering inter-related emotions that develop over time For example for abot to miss someone assumes a botrsquos happinessenjoymentexperienced during a prior interaction with a user a botrsquosability to feel longing should follow its prior display of joyshared with the user This requires critically formulating in-tentions behind communicative moves [44] of any affectivebot There are several paths for developing emotional dis-plays To list a few (1) offer one type of consistent emotionalexpressions as Vincent did (2) design emotional expressionsthat may be appropriate for different target groups in tan-dem with the implication below and (3) give users controlover how their chatbots ldquofeelrdquo towards them The caveat forthe third recommendation is that the user control over achatbotrsquos emotions may not aid a chatbotrsquos narrative andit also may not be helpful for all users the associated riskis that user-controlled emotions can render a chatbot lessrelatable More specifically the illusion that a chatbot canauthentically care for or be cared by another being requiressome level of perceived independence in how it ldquofeelsrdquo Werecommend designers to engage with the growing field ofaffective computing [41 45] and its discussion on ethics [8]If a chatbotrsquos goal is bettering usersrsquo psychological statesdesigners must ask if an affective bot delivers the intendedtreatment and what ethical boundaries there are in its dis-plays and elicitation of emotions Designers and users couldcontrol a chatbotrsquos emotions but what emotions a chatbotcan elicit in users is not always a priori foreseeable

Tailor chatbots to different target groups Evenwith one con-struct self-compassion we see a variety of ways a chatbotcan be configured To start people with low self-compassionmay benefit the most from Vincent as our exploratory anal-ysis shows This can mean more compassion focused sce-narios rather than neutral scenarios Women are noted toscore lower on self-compassion [51] yet older women experi-ence greater compassion than older men [29] Chatbots thatconsider gender age andor occupation can be potentiallyhelpful for increasing self-compassion To list a few examplesfor reincarnating Vincent a chatbot could be gendered asfemale or non-binary present a proactive version of compas-sion specified for women (see eg Neff [35] on speaking up

and protecting oneself from harm) talk about exam stresswith students or refer to stressful meetings or workplacebullying with employed individuals Rather than assumingthat one-size-fits-all or extreme personalization will workwe suggest designers to first approach targeted groups toclearly understand their needs For instance whether a self-compassion chatbot for all women is as effective or moreeffective than a more targeted chatbot eg at various levelsof intersectionality like race culture age etc should beconsidered given the time and resources that may be avail-able We recommend that based on research uncoveringpossible ways to design a chatbot that suits different needsand wants should be prioritized

Future works and limitationsOur study opened up new questions to be explored An av-enue to investigate is how a chatbotrsquos use of emotional lan-guage influences its interactants We posit that a sufferingchatbot induces less empathic distress than a suffering hu-man and whether or not this is the case needs to be furtherinvestigated especially for chatbots intended to be therapeu-tic helpers An awareness of onersquos own and othersrsquo sufferingwithout overwhelming empathic distress is suggested to bepossible through compassion [18 47] Hence disambiguat-ing compassion from empathic distress is critical in deploy-ing self-compassion chatbots as instantiations of positivecomputing [6] Different configurations of Vincent basedon peoplersquos gender age or occupation could improve theirself-compassion scores more effectively and if and in whatways this holds true warrants further research

There are limitations to consider Our effect size was lowerthan findings from the Woebot study (d = 044) which had34 participants who ldquoself-identified as experiencing symp-toms of depression and anxietyrdquo [15 p 2] and they measuredsymptoms of depression with the PHQ-9 questionnaire notself-compassion Falconer et alrsquos results on self-compassionscores after embodied VR experience also had a higher ef-fect size with the partial eta-squared of 036 (d = 1487) [14]which was based on 15 participants with depression Weworked with a general non-clinical sample and CG Vincentshowed an effect size of d = 007 (N = 34) and CR Vincentrsquoseffect size was d = 02 (N =33) Follow-up studies on self-compassion chatbots can utilize a larger sample and a moregrounded effect size One explanation for the difference ineffect size is that we did not recruit people who were clini-cally or self-proclaimed to be depressed based on the viewthat preventative mental health care can build resilience forpeople in general While Vincent andWoebot [15] share com-monalities the main measurements and targeted populationdiffered And while self-compassion was the measurementfor us and Falconer et al [14] the technology used sample

size and targeted population differed The gain andor main-tenance of healthy self-compassion as pre-emptive care maynot result in a similarly high effect size but can be psycholog-ically beneficial nonetheless More research is necessary tounderstand long-term consequences of a priori preventativecare vs a posteriori treatment of mental health woes

More broadly peoplersquos engagement with Vincent may re-flect both socially desirable reactions such as politeness to-wards machines as social actors [16 31] as well as emotionalempathy ie the ability to ldquofeel forrdquo VincentWe have not yetconcretely looked into other potential contributing factorsin bringing about self-compassion through human-chatbotinteraction Also what is difficult to gauge is the magnitudeof a chatbotrsquos perceived social and emotional complexitybased solely on messaging or text-based conversations Vin-cent lacked embodied communication which means it didnot use non-verbal modalities such as gaze voice or ges-tures that are critical in various social interactions Vincentwas a uni-modal technological entity that can be extendedthrough other complex emotional displays Thus we havenot established how people would engage with other formsof technology like robots with varying degrees and types ofembodiment alongside different combinations of modalitiesUtilizing technology appropriately for mental health carerequires many comparative renditions

5 CONCLUSIONCompassion is a key moral emotion [19] or motivation [6]that deserves to be further explored through positive comput-ing or technology for well-being Self-compassion can helppeoplersquos overall well-being [52] through kindness towardsoneself connectedness to greater humanity andmindfulnessWhile a chatbot is not a panacea for curing psychologicaldifficulties and is not meant to replace professional helpwe demonstrated that caring for a chatbot can help peoplegain greater self-compassion than being cared for by a chat-bot Our quantitative and qualitative analyses suggest thathuman-chatbot interaction is a promising arena for positivecomputing

REFERENCES[1] Anxiety and Depression Association of America (ADAA) 2016 ADAA

Reviewed Mental Health Apps httpsadaaorgmental-health-appsAccessed 2018-06-25

[2] Arthur Aron Elaine N Aron and Danny Smollan 1992 Inclusion ofOther in the Self Scale and the structure of interpersonal closenessJournal of Personality and Social Psychology 63 4 (1992) 596 httpsdoiorg1010370022-3514634596

[3] Virginia Braun and Victoria Clarke 2006 Using thematic analysisin psychology Qualitative Research in Psychology 3 2 (2006) 77ndash101httpsdoiorg1011911478088706qp063oa

[4] Scott Brave Clifford Nass and Kevin Hutchinson 2005 Computersthat care investigating the effects of orientation of emotion exhibited

Relatability leads to believability Participantsrsquo ability toextend the sense of common humanity to a chatbot touchesupon anthropomorphism CR Vincent was comforted as ifit were a human eg ldquoitrsquos human to make mistakesrdquo (CR)while its problems were addressed to its ldquochatbot worldrdquo egldquocommunicate whatrsquos going on to your fellow chatbotsrdquo (CR)For one participant even Vincentrsquos limitation of having astrict script was anthropomorphized ie ldquoVincent is like thefriend who always speaks about himself and what he haslearned or done and sometimes out of courtesy (not out ofcuriosity) asks how you are doing - but doesnrsquot listen to youranswer or what you actually have to say he just goes onwith his own thingrdquo (CG) Such attributed anthropomorphictraits depended on participantsrsquo willingness take Vincentrsquosperspective as a chatbotCR Vincentrsquos blunders were based on common human

mishaps like being late for a meeting and dealing with un-paid bills (scenarios from [13]) Yet none of our participantsquestioned whether or not a chatbot had meetings to attendor bills to pay Vincentrsquos narrative was on how a chatbotcould be late (new updates took longer than expected) orhow it could have bills (Vincent needs to pay the hostingserver) and our participants went along with imagined sce-narios Vincent faced Rather than questioning the parametersof our scenarios on realism participants thought of how tosolve Vincentrsquos problems within the parameters of a chat-botrsquos world When relevant CR Vincent played up the ironyof having human struggles as a chatbot eg ldquoall I am is lit-erally a piece of code and I failed a programming courserdquoVincent became believable because its struggles were relat-able Granting Vincent human-likness was less literal in howpeople bonded with Vincent Vincent did not try to appearhuman but it socialized with participants about its strugglesthat humans also had People related to Vincentrsquos strugglesand believed that such struggles could arise for chatbotsShared history can lead to attachment Conversations be-

tween people as well as in human-computer interactionbecome shared history over time For instance ldquo[] commu-nicating with Vincent everyday for two weeks builds somekind of habit It makes me notice its presence and absence(which might be good) I think it has a potential to be a goodcompanion and improve the mood especially if someone isfeeling lonelyrdquo (CG) Thus frequent communication with achatbot in a given duration can form expectations ldquoI reallymissed Vincent when he started our conversation laterdquo (CR)The level of attachment for some participants was higherthan others eg after the experiment we saw reactions suchas ldquocan I keep himrdquo (CG)When Vincent prepared participants for its daily good-

byes eg ldquoI have some chatbot things to do Defragmentmy server stack Buy aluminum foil to make fashionablehats withrdquo what was intended as humor can be interpreted

differently ie server defragmentation could be life-or-deathfor a chatbot Some people can be confused worried or evenangered when a chatbot they care about does not respondThus one reaction was ldquothe asshole decided to delete itsstack and when I said itrsquod die it just didnrsquot reply You canrsquotgo making people worried about a freaking chatbotrdquo (CR)People may miss a chatbot that suddenly leaves them orsincerely worry about its well-being This is the positiveand negative aspect of a relatable a chatbot some partic-ipants found common-humanity in Vincent and of thoseparticipants a few possibly related more through empathicdistress rather than through compassion If two weeks canbring about strong signs of attachment longer periods ofinteraction may heighten the level of attachment to differentdegrees and in different ways per person

Emotional reciprocity with chatbots As mentioned beforemost participants were able to respond to CR Vincentrsquos emo-tional displays on a practical level eg recommending howto fix a problem or advising Vincent on how to adjust itsemotions eg telling Vincent to stay positive To add somepeople may not appreciate chatbots demonstrating feelingsOthers may reciprocate or feel comforted by a chatbotrsquos ex-pressed emotions even if a chatbot is perceived as incapableof having emotions The more nuanced point is that Vin-centrsquos display of emotions was noted to bring conflictingfeelings For instance ldquowhen Vincent would show emotions(for example lsquolove talking to yoursquo lsquomiss yoursquo) that would feelweird because I know I am talking to a chatbot and it proba-bly is not that developed that it does have feelings But theusage of such words does feels nice compared to when a hu-man would say them So I had conflicted feelings about thesekind of expressionsrdquo (CG) The participant felt conflictedabout how to process Vincentrsquos emotional outreach

Importantly the participant suggested heshemay bemorecomfortable with a chatbot saying ldquomiss yourdquo than a humanTo conjecture the participant could mean that there was nosocial pressure due to a chatbot not expecting or needinghimher to say ldquoI miss you toordquo People often feel obligatedto respond to sincere emotions of others with similarly va-lenced emotional displays even if they do not feel the samesincere emotions towards them Such pressure may not holdfor technological entities Perhaps to miss someone impliesa certain history in a relationship so to hear that from aperson one met less than two weeks ago may feel awkwardor insincere whereas a chatbot would not be expected toknow or abide by certain social conventions If two peopleknew beforehand they will get to know each other for a maxi-mum duration of two weeks (as our participants knew beforemeeting Vincent) and never be in touch again their emo-tional performance may adjust accordingly The timescalefor intensifying socially acceptable emotional expressions

in human-chatbot interactions and human-human interac-tions may differ The ldquolifespanrdquo of a chatbot is not equatableto a personrsquos lifespan And the distinction between super-ficial vs genuine emotional displays from and to a chatbotis not entirely equatable to emotions people share and re-ciprocate between each other Currently we do not haveestablished norms on how emotions between humans andbots areshould be managed We suggest there may be dis-tinct differences compared to emotions in human-humanrelationships

Discussion and design implicationsCR Vincent adds depth to the CASA paradigm [16 31]mdash notonly do people treat a chatbot as an anthropomorphizedsocial agent but they themselves are affected by a chatbotto the extent that their self-compassion can increase whenthey are compassionate towards a chatbot Brave et alrsquosinsight on embodied conversational agents is that ldquojust aspeople respond to being cared about by other people usersrespond positively to agents that carerdquo [4 p 174] We addthat just as giving care to another human can increase onersquosself-compassion [5] caring for a chatbot can enhance onersquosown self-compassion If the dominant question has beenldquowhat can technology do for usrdquo Vincent demonstrates thatby exploring ldquowhat can we do for technologyrdquo we inad-vertently benefit from technology potentially more so thanwhen we only shape technology to serve us This claim isspecified to bots in the mental health domain and our goalwas to increase self-compassion as a target for well-being[52] rather than to reduce clinically defined symptoms ofpsychological ill-being We present our design implicationsbelow on building chatbots for psychological health carewhich primarily stem from our interpretive analysis Ourimplications are inter-related starting points that should becontextualized for each research and deployment processGive users more closed-inputs or free-input optionsMany

participants felt limited in responses they could give to Vin-cent They wanted to write to Vincent without having anypreset answers or needed more options A recommendationis to use natural language processing for a chatbot whichwill rely less on a pre-planned narrative arc and build moreon what users say This will require a longer developmentperiod The simpler option is to provide users with morefixed responses (three to four) and more opportunities foropen input

Develop a chatbotrsquos story with users Conversational agentscan be powerful storytellers [25 36] even without complexAI To deliver co-storytelling as shared history with interac-tants we suggest designers to create flexible narrative pa-rameters that people can creatively use to relate to a chatbotVincent was able to tell its story but it was less interactive inthat people could follow along with limited reaction options

due to the nature of our experiment There can be additionalcomplexities that designers can add For instance the nar-rative can take a different route depending on which closedinput options users click on We have utilized a limited num-ber of messages called ldquopathsrdquo (supplementary material) thatVincent could give depending on closed input responsesYet this practice did not change Vincentrsquos narrative Givinga chatbot ldquomemoryrdquo be it knowing basic information likenames or a more involved retention of what users say canenhance conversational storytellingTread carefully with emotional expressionsWe suggest a

broader view on what emotions are by considering inter-related emotions that develop over time For example for abot to miss someone assumes a botrsquos happinessenjoymentexperienced during a prior interaction with a user a botrsquosability to feel longing should follow its prior display of joyshared with the user This requires critically formulating in-tentions behind communicative moves [44] of any affectivebot There are several paths for developing emotional dis-plays To list a few (1) offer one type of consistent emotionalexpressions as Vincent did (2) design emotional expressionsthat may be appropriate for different target groups in tan-dem with the implication below and (3) give users controlover how their chatbots ldquofeelrdquo towards them The caveat forthe third recommendation is that the user control over achatbotrsquos emotions may not aid a chatbotrsquos narrative andit also may not be helpful for all users the associated riskis that user-controlled emotions can render a chatbot lessrelatable More specifically the illusion that a chatbot canauthentically care for or be cared by another being requiressome level of perceived independence in how it ldquofeelsrdquo Werecommend designers to engage with the growing field ofaffective computing [41 45] and its discussion on ethics [8]If a chatbotrsquos goal is bettering usersrsquo psychological statesdesigners must ask if an affective bot delivers the intendedtreatment and what ethical boundaries there are in its dis-plays and elicitation of emotions Designers and users couldcontrol a chatbotrsquos emotions but what emotions a chatbotcan elicit in users is not always a priori foreseeable

Tailor chatbots to different target groups Evenwith one con-struct self-compassion we see a variety of ways a chatbotcan be configured To start people with low self-compassionmay benefit the most from Vincent as our exploratory anal-ysis shows This can mean more compassion focused sce-narios rather than neutral scenarios Women are noted toscore lower on self-compassion [51] yet older women experi-ence greater compassion than older men [29] Chatbots thatconsider gender age andor occupation can be potentiallyhelpful for increasing self-compassion To list a few examplesfor reincarnating Vincent a chatbot could be gendered asfemale or non-binary present a proactive version of compas-sion specified for women (see eg Neff [35] on speaking up

and protecting oneself from harm) talk about exam stresswith students or refer to stressful meetings or workplacebullying with employed individuals Rather than assumingthat one-size-fits-all or extreme personalization will workwe suggest designers to first approach targeted groups toclearly understand their needs For instance whether a self-compassion chatbot for all women is as effective or moreeffective than a more targeted chatbot eg at various levelsof intersectionality like race culture age etc should beconsidered given the time and resources that may be avail-able We recommend that based on research uncoveringpossible ways to design a chatbot that suits different needsand wants should be prioritized

Future works and limitationsOur study opened up new questions to be explored An av-enue to investigate is how a chatbotrsquos use of emotional lan-guage influences its interactants We posit that a sufferingchatbot induces less empathic distress than a suffering hu-man and whether or not this is the case needs to be furtherinvestigated especially for chatbots intended to be therapeu-tic helpers An awareness of onersquos own and othersrsquo sufferingwithout overwhelming empathic distress is suggested to bepossible through compassion [18 47] Hence disambiguat-ing compassion from empathic distress is critical in deploy-ing self-compassion chatbots as instantiations of positivecomputing [6] Different configurations of Vincent basedon peoplersquos gender age or occupation could improve theirself-compassion scores more effectively and if and in whatways this holds true warrants further research

There are limitations to consider Our effect size was lowerthan findings from the Woebot study (d = 044) which had34 participants who ldquoself-identified as experiencing symp-toms of depression and anxietyrdquo [15 p 2] and they measuredsymptoms of depression with the PHQ-9 questionnaire notself-compassion Falconer et alrsquos results on self-compassionscores after embodied VR experience also had a higher ef-fect size with the partial eta-squared of 036 (d = 1487) [14]which was based on 15 participants with depression Weworked with a general non-clinical sample and CG Vincentshowed an effect size of d = 007 (N = 34) and CR Vincentrsquoseffect size was d = 02 (N =33) Follow-up studies on self-compassion chatbots can utilize a larger sample and a moregrounded effect size One explanation for the difference ineffect size is that we did not recruit people who were clini-cally or self-proclaimed to be depressed based on the viewthat preventative mental health care can build resilience forpeople in general While Vincent andWoebot [15] share com-monalities the main measurements and targeted populationdiffered And while self-compassion was the measurementfor us and Falconer et al [14] the technology used sample

size and targeted population differed The gain andor main-tenance of healthy self-compassion as pre-emptive care maynot result in a similarly high effect size but can be psycholog-ically beneficial nonetheless More research is necessary tounderstand long-term consequences of a priori preventativecare vs a posteriori treatment of mental health woes

More broadly peoplersquos engagement with Vincent may re-flect both socially desirable reactions such as politeness to-wards machines as social actors [16 31] as well as emotionalempathy ie the ability to ldquofeel forrdquo VincentWe have not yetconcretely looked into other potential contributing factorsin bringing about self-compassion through human-chatbotinteraction Also what is difficult to gauge is the magnitudeof a chatbotrsquos perceived social and emotional complexitybased solely on messaging or text-based conversations Vin-cent lacked embodied communication which means it didnot use non-verbal modalities such as gaze voice or ges-tures that are critical in various social interactions Vincentwas a uni-modal technological entity that can be extendedthrough other complex emotional displays Thus we havenot established how people would engage with other formsof technology like robots with varying degrees and types ofembodiment alongside different combinations of modalitiesUtilizing technology appropriately for mental health carerequires many comparative renditions

5 CONCLUSIONCompassion is a key moral emotion [19] or motivation [6]that deserves to be further explored through positive comput-ing or technology for well-being Self-compassion can helppeoplersquos overall well-being [52] through kindness towardsoneself connectedness to greater humanity andmindfulnessWhile a chatbot is not a panacea for curing psychologicaldifficulties and is not meant to replace professional helpwe demonstrated that caring for a chatbot can help peoplegain greater self-compassion than being cared for by a chat-bot Our quantitative and qualitative analyses suggest thathuman-chatbot interaction is a promising arena for positivecomputing

REFERENCES[1] Anxiety and Depression Association of America (ADAA) 2016 ADAA

Reviewed Mental Health Apps httpsadaaorgmental-health-appsAccessed 2018-06-25

[2] Arthur Aron Elaine N Aron and Danny Smollan 1992 Inclusion ofOther in the Self Scale and the structure of interpersonal closenessJournal of Personality and Social Psychology 63 4 (1992) 596 httpsdoiorg1010370022-3514634596

[3] Virginia Braun and Victoria Clarke 2006 Using thematic analysisin psychology Qualitative Research in Psychology 3 2 (2006) 77ndash101httpsdoiorg1011911478088706qp063oa

[4] Scott Brave Clifford Nass and Kevin Hutchinson 2005 Computersthat care investigating the effects of orientation of emotion exhibited

in human-chatbot interactions and human-human interac-tions may differ The ldquolifespanrdquo of a chatbot is not equatableto a personrsquos lifespan And the distinction between super-ficial vs genuine emotional displays from and to a chatbotis not entirely equatable to emotions people share and re-ciprocate between each other Currently we do not haveestablished norms on how emotions between humans andbots areshould be managed We suggest there may be dis-tinct differences compared to emotions in human-humanrelationships

Discussion and design implicationsCR Vincent adds depth to the CASA paradigm [16 31]mdash notonly do people treat a chatbot as an anthropomorphizedsocial agent but they themselves are affected by a chatbotto the extent that their self-compassion can increase whenthey are compassionate towards a chatbot Brave et alrsquosinsight on embodied conversational agents is that ldquojust aspeople respond to being cared about by other people usersrespond positively to agents that carerdquo [4 p 174] We addthat just as giving care to another human can increase onersquosself-compassion [5] caring for a chatbot can enhance onersquosown self-compassion If the dominant question has beenldquowhat can technology do for usrdquo Vincent demonstrates thatby exploring ldquowhat can we do for technologyrdquo we inad-vertently benefit from technology potentially more so thanwhen we only shape technology to serve us This claim isspecified to bots in the mental health domain and our goalwas to increase self-compassion as a target for well-being[52] rather than to reduce clinically defined symptoms ofpsychological ill-being We present our design implicationsbelow on building chatbots for psychological health carewhich primarily stem from our interpretive analysis Ourimplications are inter-related starting points that should becontextualized for each research and deployment processGive users more closed-inputs or free-input optionsMany

participants felt limited in responses they could give to Vin-cent They wanted to write to Vincent without having anypreset answers or needed more options A recommendationis to use natural language processing for a chatbot whichwill rely less on a pre-planned narrative arc and build moreon what users say This will require a longer developmentperiod The simpler option is to provide users with morefixed responses (three to four) and more opportunities foropen input

Develop a chatbotrsquos story with users Conversational agentscan be powerful storytellers [25 36] even without complexAI To deliver co-storytelling as shared history with interac-tants we suggest designers to create flexible narrative pa-rameters that people can creatively use to relate to a chatbotVincent was able to tell its story but it was less interactive inthat people could follow along with limited reaction options

due to the nature of our experiment There can be additionalcomplexities that designers can add For instance the nar-rative can take a different route depending on which closedinput options users click on We have utilized a limited num-ber of messages called ldquopathsrdquo (supplementary material) thatVincent could give depending on closed input responsesYet this practice did not change Vincentrsquos narrative Givinga chatbot ldquomemoryrdquo be it knowing basic information likenames or a more involved retention of what users say canenhance conversational storytellingTread carefully with emotional expressionsWe suggest a

broader view on what emotions are by considering inter-related emotions that develop over time For example for abot to miss someone assumes a botrsquos happinessenjoymentexperienced during a prior interaction with a user a botrsquosability to feel longing should follow its prior display of joyshared with the user This requires critically formulating in-tentions behind communicative moves [44] of any affectivebot There are several paths for developing emotional dis-plays To list a few (1) offer one type of consistent emotionalexpressions as Vincent did (2) design emotional expressionsthat may be appropriate for different target groups in tan-dem with the implication below and (3) give users controlover how their chatbots ldquofeelrdquo towards them The caveat forthe third recommendation is that the user control over achatbotrsquos emotions may not aid a chatbotrsquos narrative andit also may not be helpful for all users the associated riskis that user-controlled emotions can render a chatbot lessrelatable More specifically the illusion that a chatbot canauthentically care for or be cared by another being requiressome level of perceived independence in how it ldquofeelsrdquo Werecommend designers to engage with the growing field ofaffective computing [41 45] and its discussion on ethics [8]If a chatbotrsquos goal is bettering usersrsquo psychological statesdesigners must ask if an affective bot delivers the intendedtreatment and what ethical boundaries there are in its dis-plays and elicitation of emotions Designers and users couldcontrol a chatbotrsquos emotions but what emotions a chatbotcan elicit in users is not always a priori foreseeable

Tailor chatbots to different target groups Evenwith one con-struct self-compassion we see a variety of ways a chatbotcan be configured To start people with low self-compassionmay benefit the most from Vincent as our exploratory anal-ysis shows This can mean more compassion focused sce-narios rather than neutral scenarios Women are noted toscore lower on self-compassion [51] yet older women experi-ence greater compassion than older men [29] Chatbots thatconsider gender age andor occupation can be potentiallyhelpful for increasing self-compassion To list a few examplesfor reincarnating Vincent a chatbot could be gendered asfemale or non-binary present a proactive version of compas-sion specified for women (see eg Neff [35] on speaking up

and protecting oneself from harm) talk about exam stresswith students or refer to stressful meetings or workplacebullying with employed individuals Rather than assumingthat one-size-fits-all or extreme personalization will workwe suggest designers to first approach targeted groups toclearly understand their needs For instance whether a self-compassion chatbot for all women is as effective or moreeffective than a more targeted chatbot eg at various levelsof intersectionality like race culture age etc should beconsidered given the time and resources that may be avail-able We recommend that based on research uncoveringpossible ways to design a chatbot that suits different needsand wants should be prioritized

Future works and limitationsOur study opened up new questions to be explored An av-enue to investigate is how a chatbotrsquos use of emotional lan-guage influences its interactants We posit that a sufferingchatbot induces less empathic distress than a suffering hu-man and whether or not this is the case needs to be furtherinvestigated especially for chatbots intended to be therapeu-tic helpers An awareness of onersquos own and othersrsquo sufferingwithout overwhelming empathic distress is suggested to bepossible through compassion [18 47] Hence disambiguat-ing compassion from empathic distress is critical in deploy-ing self-compassion chatbots as instantiations of positivecomputing [6] Different configurations of Vincent basedon peoplersquos gender age or occupation could improve theirself-compassion scores more effectively and if and in whatways this holds true warrants further research

There are limitations to consider Our effect size was lowerthan findings from the Woebot study (d = 044) which had34 participants who ldquoself-identified as experiencing symp-toms of depression and anxietyrdquo [15 p 2] and they measuredsymptoms of depression with the PHQ-9 questionnaire notself-compassion Falconer et alrsquos results on self-compassionscores after embodied VR experience also had a higher ef-fect size with the partial eta-squared of 036 (d = 1487) [14]which was based on 15 participants with depression Weworked with a general non-clinical sample and CG Vincentshowed an effect size of d = 007 (N = 34) and CR Vincentrsquoseffect size was d = 02 (N =33) Follow-up studies on self-compassion chatbots can utilize a larger sample and a moregrounded effect size One explanation for the difference ineffect size is that we did not recruit people who were clini-cally or self-proclaimed to be depressed based on the viewthat preventative mental health care can build resilience forpeople in general While Vincent andWoebot [15] share com-monalities the main measurements and targeted populationdiffered And while self-compassion was the measurementfor us and Falconer et al [14] the technology used sample

size and targeted population differed The gain andor main-tenance of healthy self-compassion as pre-emptive care maynot result in a similarly high effect size but can be psycholog-ically beneficial nonetheless More research is necessary tounderstand long-term consequences of a priori preventativecare vs a posteriori treatment of mental health woes

More broadly peoplersquos engagement with Vincent may re-flect both socially desirable reactions such as politeness to-wards machines as social actors [16 31] as well as emotionalempathy ie the ability to ldquofeel forrdquo VincentWe have not yetconcretely looked into other potential contributing factorsin bringing about self-compassion through human-chatbotinteraction Also what is difficult to gauge is the magnitudeof a chatbotrsquos perceived social and emotional complexitybased solely on messaging or text-based conversations Vin-cent lacked embodied communication which means it didnot use non-verbal modalities such as gaze voice or ges-tures that are critical in various social interactions Vincentwas a uni-modal technological entity that can be extendedthrough other complex emotional displays Thus we havenot established how people would engage with other formsof technology like robots with varying degrees and types ofembodiment alongside different combinations of modalitiesUtilizing technology appropriately for mental health carerequires many comparative renditions

5 CONCLUSIONCompassion is a key moral emotion [19] or motivation [6]that deserves to be further explored through positive comput-ing or technology for well-being Self-compassion can helppeoplersquos overall well-being [52] through kindness towardsoneself connectedness to greater humanity andmindfulnessWhile a chatbot is not a panacea for curing psychologicaldifficulties and is not meant to replace professional helpwe demonstrated that caring for a chatbot can help peoplegain greater self-compassion than being cared for by a chat-bot Our quantitative and qualitative analyses suggest thathuman-chatbot interaction is a promising arena for positivecomputing

REFERENCES[1] Anxiety and Depression Association of America (ADAA) 2016 ADAA

Reviewed Mental Health Apps httpsadaaorgmental-health-appsAccessed 2018-06-25

[2] Arthur Aron Elaine N Aron and Danny Smollan 1992 Inclusion ofOther in the Self Scale and the structure of interpersonal closenessJournal of Personality and Social Psychology 63 4 (1992) 596 httpsdoiorg1010370022-3514634596

[3] Virginia Braun and Victoria Clarke 2006 Using thematic analysisin psychology Qualitative Research in Psychology 3 2 (2006) 77ndash101httpsdoiorg1011911478088706qp063oa

[4] Scott Brave Clifford Nass and Kevin Hutchinson 2005 Computersthat care investigating the effects of orientation of emotion exhibited

and protecting oneself from harm) talk about exam stresswith students or refer to stressful meetings or workplacebullying with employed individuals Rather than assumingthat one-size-fits-all or extreme personalization will workwe suggest designers to first approach targeted groups toclearly understand their needs For instance whether a self-compassion chatbot for all women is as effective or moreeffective than a more targeted chatbot eg at various levelsof intersectionality like race culture age etc should beconsidered given the time and resources that may be avail-able We recommend that based on research uncoveringpossible ways to design a chatbot that suits different needsand wants should be prioritized

Future works and limitationsOur study opened up new questions to be explored An av-enue to investigate is how a chatbotrsquos use of emotional lan-guage influences its interactants We posit that a sufferingchatbot induces less empathic distress than a suffering hu-man and whether or not this is the case needs to be furtherinvestigated especially for chatbots intended to be therapeu-tic helpers An awareness of onersquos own and othersrsquo sufferingwithout overwhelming empathic distress is suggested to bepossible through compassion [18 47] Hence disambiguat-ing compassion from empathic distress is critical in deploy-ing self-compassion chatbots as instantiations of positivecomputing [6] Different configurations of Vincent basedon peoplersquos gender age or occupation could improve theirself-compassion scores more effectively and if and in whatways this holds true warrants further research

There are limitations to consider Our effect size was lowerthan findings from the Woebot study (d = 044) which had34 participants who ldquoself-identified as experiencing symp-toms of depression and anxietyrdquo [15 p 2] and they measuredsymptoms of depression with the PHQ-9 questionnaire notself-compassion Falconer et alrsquos results on self-compassionscores after embodied VR experience also had a higher ef-fect size with the partial eta-squared of 036 (d = 1487) [14]which was based on 15 participants with depression Weworked with a general non-clinical sample and CG Vincentshowed an effect size of d = 007 (N = 34) and CR Vincentrsquoseffect size was d = 02 (N =33) Follow-up studies on self-compassion chatbots can utilize a larger sample and a moregrounded effect size One explanation for the difference ineffect size is that we did not recruit people who were clini-cally or self-proclaimed to be depressed based on the viewthat preventative mental health care can build resilience forpeople in general While Vincent andWoebot [15] share com-monalities the main measurements and targeted populationdiffered And while self-compassion was the measurementfor us and Falconer et al [14] the technology used sample

size and targeted population differed The gain andor main-tenance of healthy self-compassion as pre-emptive care maynot result in a similarly high effect size but can be psycholog-ically beneficial nonetheless More research is necessary tounderstand long-term consequences of a priori preventativecare vs a posteriori treatment of mental health woes

More broadly peoplersquos engagement with Vincent may re-flect both socially desirable reactions such as politeness to-wards machines as social actors [16 31] as well as emotionalempathy ie the ability to ldquofeel forrdquo VincentWe have not yetconcretely looked into other potential contributing factorsin bringing about self-compassion through human-chatbotinteraction Also what is difficult to gauge is the magnitudeof a chatbotrsquos perceived social and emotional complexitybased solely on messaging or text-based conversations Vin-cent lacked embodied communication which means it didnot use non-verbal modalities such as gaze voice or ges-tures that are critical in various social interactions Vincentwas a uni-modal technological entity that can be extendedthrough other complex emotional displays Thus we havenot established how people would engage with other formsof technology like robots with varying degrees and types ofembodiment alongside different combinations of modalitiesUtilizing technology appropriately for mental health carerequires many comparative renditions

5 CONCLUSIONCompassion is a key moral emotion [19] or motivation [6]that deserves to be further explored through positive comput-ing or technology for well-being Self-compassion can helppeoplersquos overall well-being [52] through kindness towardsoneself connectedness to greater humanity andmindfulnessWhile a chatbot is not a panacea for curing psychologicaldifficulties and is not meant to replace professional helpwe demonstrated that caring for a chatbot can help peoplegain greater self-compassion than being cared for by a chat-bot Our quantitative and qualitative analyses suggest thathuman-chatbot interaction is a promising arena for positivecomputing

REFERENCES[1] Anxiety and Depression Association of America (ADAA) 2016 ADAA

Reviewed Mental Health Apps httpsadaaorgmental-health-appsAccessed 2018-06-25

[2] Arthur Aron Elaine N Aron and Danny Smollan 1992 Inclusion ofOther in the Self Scale and the structure of interpersonal closenessJournal of Personality and Social Psychology 63 4 (1992) 596 httpsdoiorg1010370022-3514634596

[3] Virginia Braun and Victoria Clarke 2006 Using thematic analysisin psychology Qualitative Research in Psychology 3 2 (2006) 77ndash101httpsdoiorg1011911478088706qp063oa

[4] Scott Brave Clifford Nass and Kevin Hutchinson 2005 Computersthat care investigating the effects of orientation of emotion exhibited

by an embodied computer agent International Journal of Human-computer Studies 62 2 (2005) 161ndash178 httpsdoiorg101016jijhcs200411002

[5] Juliana G Breines and Serena Chen 2013 Activating the inner care-giver The role of support-giving schemas in increasing state self-compassion Journal of Experimental Social Psychology 49 1 (2013)58ndash64 httpsdoiorg101016jjesp201207015

[6] Rafael A Calvo and Dorian Peters 2014 Positive Computing Technologyfor Wellbeing and Human Potential MIT Press

[7] Hyun-Chul Cho and Shuzo Abe 2013 Is two-tailed testing for di-rectional research hypotheses tests legitimate Journal of BusinessResearch 66 9 (2013) 1261ndash1266 httpsdoiorg101016jjbusres201202023

[8] Roddy Cowie 2015 Ethical issues in affective computing In The OxfordHandbook of Affective Computing in Calvo R A DrsquoMello S GratchJ and Kappas A (Eds) Oxford Library of Psychology 334ndash338

[9] Nils Dahlbaumlck Arne Joumlnsson and Lars Ahrenberg 1993 Wizard of Ozstudies - why and how Knowledge-based Systems 6 4 (1993) 258ndash266httpsdoiorg101145169891169968

[10] Robert Dale 2016 The return of the chatbots Natural Lan-guage Engineering 22 5 (2016) 811ndash817 httpsdoiorg101017S1351324916000243

[11] Kerstin Dautenhahn 2007 Socially intelligent robots Dimensionsof humanndashrobot interaction Philosophical Transactions of the RoyalSociety of London B Biological Sciences 362 1480 (2007) 679ndash704 httpsdoiorg101098rstb20062004

[12] Rachel Dodge Annette P Daly Jan Huyton and Lalage D Sanders2012 The challenge of defining wellbeing International Journal ofWellbeing 2 3 (2012) 222ndash235 httpsdoiorg105502ijwv2i34

[13] Caroline J Falconer John A King and Chris R Brewin 2015Demonstrating mood repair with a situation-based measure of self-compassion and self-criticism Psychology and Psychotherapy TheoryResearch and Practice 88 4 (2015) 351ndash365 httpsdoiorg101111papt12056

[14] Caroline J Falconer Aitor Rovira John A King Paul Gilbert AngusAntley Pasco Fearon Neil Ralph Mel Slater and Chris R Brewin 2016Embodying self-compassion within virtual reality and its effects onpatients with depression BJPsych Open 2 1 (2016) 74ndash80

[15] Kathleen Kara Fitzpatrick Alison Darcy and Molly Vierhile 2017Delivering cognitive behavior therapy to young adults with symptomsof depression and anxiety using a fully automated conversational agent(Woebot) a randomized controlled trial JMIR Mental Health 4 2 (2017)httpsdoiorg102196mental7785

[16] BJ Fogg and Clifford Nass 1997 How users reciprocate to computersan experiment that demonstrates behavior change In CHIrsquo97 ExtendedAbstracts on Human Factors in Computing Systems (CHI EA rsquo97) ACMNew York NY USA 331ndash332 httpsdoiorg10114511202121120419

[17] Christopher K Germer and Kristin D Neff 2013 Self-compassion inclinical practice Journal of Clinical Psychology 69 8 (2013) 856ndash867httpsdoiorg101002jclp22021

[18] Paul Gilbert 2014 The origins and nature of compassion focusedtherapy British Journal of Clinical Psychology 53 1 (2014) 6ndash41

[19] Jonathan Haidt 2003 The moral emotions In Handbook of AffectiveSciences Vol 11 R J Davidson K Scherer and H H Goldsmith (Eds)Oxford University Press 852ndash870

[20] Jeffrey T Hancock Christopher Landrigan and Courtney Silver 2007Expressing emotion in text-based communication In Proceedings of the2007 CHI Conference on Human Factors in Computing Systems (CHI rsquo07)ACM New York NY USA 929ndash932 httpsdoiorg10114512406241240764

[21] Annika Howells Itai Ivtzan and Francisco Jose Eiroa-Orosa 2016Putting the rsquoapprsquo in happiness A randomised controlled trial of a

smartphone-based mindfulness intervention to enhance wellbeingJournal of Happiness Studies 17 1 (2016) 163ndash185 httpsdoiorg101007s10902-014-9589-1

[22] Ichiro Kawachi and Lisa F Berkman 2001 Social ties and mental healthJournal of Urban Health 78 3 (2001) 458ndash467 httpsdoiorg101093jurban783458

[23] Sara H Konrath Edward H OrsquoBrien and Courtney Hsing 2011Changes in dispositional empathy in American college students overtime A meta-analysis Personality and Social Psychology Review 15 2(2011) 180ndash198

[24] Minha Lee Lily Frank Femke Beute Yvonne de Kort and WijnandIJsselsteijn 2017 Bots mind the social-technical gap In Proceedings of15th European Conference on Computer-Supported Cooperative Work-Exploratory Papers European Society for Socially Embedded Technolo-gies (EUSSET) httpsdoiorg1018420ecscw2017-14

[25] Michael Mateas and Phoebe Sengers 1999 Narrative Intelligence InNarrative Intelligence Papers from the AAAI Fall Symposium (1999)AAAI TR FS-99-01 AAAI Menlo Park CA

[26] Shizuko Matsuzoe and Fumihide Tanaka 2012 How smartly shouldrobots behave Comparative investigation on the learning ability ofa care-receiving robot In Proceedings of the 21th IEEE InternationalWorkshop on Robot and Human Interactive Communication (RO-MANrsquo12) IEEE 339ndash344 httpsdoiorg101109ROMAN20126343776

[27] EmilyMcRae 2012 The Psychology of Moral Judgment and Perceptionin Indo-Tibetan Buddhist Ethics In The Oxford Handbook of BuddhistEthics in Cozort D and Shields JM (Eds) Oxford University Press335ndash358

[28] Youngme Moon and Clifford Nass 1996 How ldquorealrdquo are computerpersonalities Psychological responses to personality types in human-computer interaction Communication Research 23 6 (1996) 651ndash674httpsdoiorg101177009365096023006002

[29] Raeanne C Moore Arsquoverria Sirkin Martin Allison R Kaup Wesley KThompson Matthew E Peters Dilip V Jeste Shahrokh Golshan andLisa T Eyler 2015 From suffering to caring a model of differencesamong older adults in levels of compassion International Journal ofGeriatric Psychiatry 30 2 (2015) 185ndash191 httpsdoiorg101002gps4123

[30] David G Myers and Ed Diener 1995 Who is happy PsychologicalScience 6 1 (1995) 10ndash19 httpsdoiorg101111j1467-92801995tb00298x

[31] Clifford Nass Jonathan Steuer and Ellen R Tauber 1994 Computersare social actors In Proceedings of the SIGCHI Conference on HumanFactors in Computing Systems (CHI rsquo94) ACM New York NY USA72ndash78

[32] Kristin Neff 2003 Self-compassion An alternative conceptualizationof a healthy attitude toward oneself Self and Identity 2 2 (2003)85ndash101 httpsdoiorg10108015298860309032

[33] Kristin Neff 2003 Test how self-compassionate you are httpsself-compassionorgtest-how-self-compassionate-you-are Ac-cessed 2018-07-01

[34] Kristin Neff 2008 Self-Compassion Exercises httpself-compassionorgcategoryexercisesexercises Accessed 2018-07-01

[35] Kristin Neff 2018 Why Women Need Fierce Self-Compassionhttpsgreatergoodberkeleyeduarticleitemwhy_women_need_fierce_self_compassion Accessed 2018-12-30

[36] Chrystopher L Nehaniv 1999 Narrative for artifacts Transcendingcontext and self In Narrative Intelligence Papers from the AAAI FallSymposium (1999) AAAI TR FS-99-01 AAAI Menlo Park CA

[37] Shaun Nichols 2004 Sentimental Rules On the Natural Foundations ofMoral Judgment Oxford University Press

[38] Amy Ellis Nutt 2017 The Washington Post ldquoThe Woe-bot will see you nowrdquo - the rise of chatbot therapy

httpswwwwashingtonpostcomnewsto-your-healthwp20171203the-woebot-will-see-you-now-the-rise-of-chatbot-therapyAccessed 2018-07-16

[39] World Health Organization 2017 Mental Health ATLAS2017 httpappswhointirisbitstreamhandle106652727359789241514019-engpdf Accessed 2018-06-27

[40] World Health Organization 2017 Mental health massive scale-upof resources needed if global targets are to be met httpwwwwhointmental_healthevidenceatlasatlas_2017_web_noteen Accessed2018-06-27

[41] Rosalind W Picard 2003 Affective computing Challenges Inter-national Journal of Human-Computer Studies 59 1-2 (2003) 55ndash64httpsdoiorg101016S1071-5819(03)00052-1

[42] Gene Reeves 2012 A Perspective on Ethics in the Lotus Sutra In TheOxford Handbook of Buddhist Ethics in Cozort D and Shields JM(Eds) Oxford University Press 335ndash358

[43] David Rodrigues Mariacutelia Prada Rui Gaspar Margarida V Garridoand Diniz Lopes 2018 Lisbon Emoji and Emoticon Database (LEED)Norms for emoji and emoticons in seven evaluative dimensions Be-havior Research Methods 50 1 (2018) 392ndash405 httpsdoiorg103758s13428-017-0878-6

[44] Andrea Scarantino 2017 How to Do Things with Emotional Expres-sions The Theory of Affective Pragmatics Psychological Inquiry 282-3 (2017) 165ndash185 httpsdoiorg1010801047840X20171328951

[45] Klaus R Scherer Tanja Baumlnziger and Etienne Roesch 2010 A Blueprintfor Affective Computing A Sourcebook and Manual Oxford University

Press[46] Arthur Schopenhauer 1995 On the Basis of Morality Hackett Publish-

ing[47] Acharya Shantideva 1979 Guide to the Bodhisattvarsquos Way of Life

Batchelor S trans Dharamsala India Library of Tibetan Works andArchives

[48] Jonathan A Smith 1996 Beyond the divide between cognition anddiscourse Using interpretative phenomenological analysis in healthpsychology Psychology and Health 11 2 (1996) 261ndash271 httpsdoiorg10108008870449608400256

[49] Fumihide Tanaka and Shizuko Matsuzoe 2012 Children teach a care-receiving robot to promote their learning Field experiments in a class-room for vocabulary learning Journal of Human-Robot Interaction 1 1(2012) 78ndash95 httpsdoiorg105898JHRI11Tanaka

[50] Joseph Weizenbaum 1966 ELIZA - a computer program for the studyof natural language communication between man and machine Com-mun ACM 9 1 (1966) 36ndash45 httpsdoiorg101145365153365168

[51] Lisa M Yarnell Rose E Stafford Kristin D Neff Erin D Reilly Marissa CKnox and Michael Mullarkey 2015 Meta-analysis of gender dif-ferences in self-compassion Self and Identity 14 5 (2015) 499ndash520httpsdoiorg1010801529886820151029966

[52] Ulli Zessin Oliver Dickhaumluser and Sven Garbade 2015 The rela-tionship between self-compassion and well-being A meta-analysisApplied Psychology Health and Well-Being 7 3 (2015) 340ndash364 httpsdoiorg101111aphw12051