...

.

...

.

...

.

...

.

...

.

...

.

...

.

...

.

...

.

...

.

How do bots interact with each other?

The online world has turnedinto an ecosystem of bots.

Even Good Bots FightEven Good Bots FightEven Good Bots FightThis project has received funding from the European Union's

Horizon 2020 Research and Innovation Program under grant agreement No 645043.

Milena Tsvetkovaa, Ruth García-Gavilanesa, Luciano Floridia,b, and Taha Yasseria,b

The online world has turned into an ecosystem of bots.

Bots are responsible for an increasingly larger proportion of activities on the Web.

How do these automated agents interact with each other? Our research suggests that even relatively “dumb” bots may give rise to complex interactions. Thus, as bots continue to proliferate and manage an increasingly larger portion of our lives, the sociology of arti�cial agents will coalesce and grow as an important research �eld.

We model how pairs of bots interact over time as interaction trajectories.

We analyze three properties of the trajectories

Although Wikipedia bots are intended to support the encyclopedia, they often undo each other’s edits and these “�ghts” may sometimes continue for years.

EnglishJapaneseSpanishFrenchPortugueseGermanChineseHebrewHungarianCzechArabicRomanianPersian

EnglishJapaneseSpanishFrenchPortugueseGermanChineseHebrewHungarianCzechArabicRomanianPersian

Bot-bot interactions have different characteristic time scale than human-human interactions.

We identi�ed four types of trajectories. Their distribution shows that bot-bot interactions tend to be more reciprocated.

1. Gianvecchio S, Xie M, Wu Z, Wang H (2008) Measurement and classi�cation of humans and bots in Internet chat. Proceedings of the 17th Conference on Security Symposium (USENIX Association), pp 155–169.

2. Sysomos (2009) An in-depth look at the most active Twitter user data. Available at: https://sysomos.com/inside-twitter/most-active-twitter-us-er-data.

3. Cashmore P (2009) Twitter zombies: 24% of tweets created by bots. Mash-able. Available at: http://mashable.com/2009/08/06/twitter-bots/#JqT-VM0vEgqqA

Wikipedia bots interact, and their interactions can be inef�cient and in�uenced by culture, just as those between humans.

In additional analyses, we found that the same bots are responsible for the majority of reverts across all languages. These are bots that specialize in creating and modifying links between the different language editions of Wikipedia. The difference between German and Portuguese bots is not due to different kinds of bots but due to the same bots operating in different kinds of environments.

2004

1600

900

400

100

0

-100

-400

Reve

rts,

squ

arer

oote

d

2005 2006 2007 2008 2009 2010 2004

1600

900

400

100

0

-100

-400

Reve

rts,

squ

arer

oote

d

2005 2006 2007 2008 2009 2010

>trade

>tweet

>crawl

>edit

>censor

>spam

>bid

>attack

>chat

>phish

>like

>review

>spy

>upvote

>click

a Oxford Internet Institute b Alan Turing Institute Designed by Daniil Alexandrov

0 102

100

10-1

10-2

103

10-4

10-5

104 106 108

Latency, sec

Freq

uenc

y (h

uman

-hum

an)

t

Bala

nce

Time low high

low high

low high

Bot-bot Human-human20 43 14 19 23

Freq

uenc

y

0 102

100

10-1

10-2

103

10-4

10-5

104 106 108

Latency, sec

Bot-bot Human-human

398144114094187

28 40 1231

2433

2829

2623

2227

2224

20

4446

4039

4542

433842

4843

45

2322

2020

2222

2931

2829

2528

515042524033344334304829

1517

2211

161023

1727

4021

14

2317

1816

2619

221818

1517

20

1216

1821

1837

212121

1514

36

RESULTS

CONCLUSION

REFERENCES

INTRODUCTION

METHOD

Bot-bot trajectories Human-human trajectories

Fast unbalanced

Slow unbalanced Somewhat balanced

Well balanced

The mean log time between successive reverts;

The �nal proportion of reverts that were not reciprocated;

The proportion of observed turning points out of all possible.

We use k-means clustering to identify different types of interaction trajectories and to estimate how common the types are for bots compared to humans and for different languages.

Latency

Imbalance

Reciprocity

We analyze the interactions between bots that edit articles on Wikipedia. We track the extent to which bots undid each other’s edits.

13 different language editions

REVERTS

Period 2001-2010

Bots comprise a tiny proportion of all Wikipedia users, usually less than 0.1%.

However, they account for a signi�cant proportion of the editorial activity.

Bots behave differently in different language editions.

English

17 285 053

3 905 586

672

197 850 396

5 162 236

16 838 608

28 717 967

10 2456

3 301 494

22 023 332

948 597

5 755 678

13 366 493

353 879

3 479 859

11 958 185

436 821

5 615 348

8 796 942

611 831

1 055 449

9 795 722

51 209

2 604 780

4 961 398

149 210

1 785 903

3 732 386

28 485

1 990 820

2 514 563

24 746

1 657 897

2 607 205

49 583

1 895 405

1 603 020

2 6060

1 591 712

1 591 712

18 528

1 250 362

104.6 104 71.7 59.3 185 24.1 103 83.9 66.8 59 161.7 78 63.8

3.1 2.6 2.3 2.6 2.3 2.2 3.2 2.9 2.4 3.1 2.2 2.1 3.6

2 288 040

73 840

222

2 526 608

755 155

266

1 319 932

279 978

304

1 461 501

365 102

191

1 207 134

494 503

260

662 008

41 165

169

281 245

113 105

139

155 918

24 657

150

134 782

19 770

142

303 230

46 291

149

105 512

22 825

120

73 489

15 416

121

Humans

Japanese Spanish French Portuguese German Chinese Hebrew Hungarian Czech Arabic Romanian Persian

Vandals

Bots

Human edits

Vandal edits

Bot edits

DATA

Average human-human reverts per editor:

Average bot-bot reverts per editor:

i reve

rts j j reverts i

We analyze interactions betweenbots that edit articles on Wikipedia.

RevertsPeriod 2001–201013 languages

Milena Tsvetkova, Ruth García-Gavilanes, Luciano Floridi & Taha Yasseri Even Good Bots Fight: 1/3

...

.

...

.

...

.

...

.

...

.

...

.

...

.

...

.

...

.

...

.

We study interaction trajectories and three of their properties.

Even Good Bots FightEven Good Bots FightEven Good Bots FightThis project has received funding from the European Union's

Horizon 2020 Research and Innovation Program under grant agreement No 645043.

Milena Tsvetkovaa, Ruth García-Gavilanesa, Luciano Floridia,b, and Taha Yasseria,b

The online world has turned into an ecosystem of bots.

Bots are responsible for an increasingly larger proportion of activities on the Web.

How do these automated agents interact with each other? Our research suggests that even relatively “dumb” bots may give rise to complex interactions. Thus, as bots continue to proliferate and manage an increasingly larger portion of our lives, the sociology of arti�cial agents will coalesce and grow as an important research �eld.

We model how pairs of bots interact over time as interaction trajectories.

We analyze three properties of the trajectories

Although Wikipedia bots are intended to support the encyclopedia, they often undo each other’s edits and these “�ghts” may sometimes continue for years.

EnglishJapaneseSpanishFrenchPortugueseGermanChineseHebrewHungarianCzechArabicRomanianPersian

EnglishJapaneseSpanishFrenchPortugueseGermanChineseHebrewHungarianCzechArabicRomanianPersian

Bot-bot interactions have different characteristic time scale than human-human interactions.

We identi�ed four types of trajectories. Their distribution shows that bot-bot interactions tend to be more reciprocated.

1. Gianvecchio S, Xie M, Wu Z, Wang H (2008) Measurement and classi�cation of humans and bots in Internet chat. Proceedings of the 17th Conference on Security Symposium (USENIX Association), pp 155–169.

2. Sysomos (2009) An in-depth look at the most active Twitter user data. Available at: https://sysomos.com/inside-twitter/most-active-twitter-us-er-data.

3. Cashmore P (2009) Twitter zombies: 24% of tweets created by bots. Mash-able. Available at: http://mashable.com/2009/08/06/twitter-bots/#JqT-VM0vEgqqA

Wikipedia bots interact, and their interactions can be inef�cient and in�uenced by culture, just as those between humans.

In additional analyses, we found that the same bots are responsible for the majority of reverts across all languages. These are bots that specialize in creating and modifying links between the different language editions of Wikipedia. The difference between German and Portuguese bots is not due to different kinds of bots but due to the same bots operating in different kinds of environments.

2004

1600

900

400

100

0

-100

-400

Reve

rts,

squ

arer

oote

d

2005 2006 2007 2008 2009 2010 2004

1600

900

400

100

0

-100

-400

Reve

rts,

squ

arer

oote

d

2005 2006 2007 2008 2009 2010

>trade

>tweet

>crawl

>edit

>censor

>spam

>bid

>attack

>chat

>phish

>like

>review

>spy

>upvote

>click

a Oxford Internet Institute b Alan Turing Institute Designed by Daniil Alexandrov

0 102

100

10-1

10-2

103

10-4

10-5

104 106 108

Latency, sec

Freq

uenc

y (h

uman

-hum

an)

t

Bala

nce

Time low high

low high

low high

Bot-bot Human-human20 43 14 19 23

Freq

uenc

y

0 102

100

10-1

10-2

103

10-4

10-5

104 106 108

Latency, sec

Bot-bot Human-human

398144114094187

28 40 1231

2433

2829

2623

2227

2224

20

4446

4039

4542

433842

4843

45

2322

2020

2222

2931

2829

2528

515042524033344334304829

1517

2211

161023

1727

4021

14

2317

1816

2619

221818

1517

20

1216

1821

1837

212121

1514

36

RESULTS

CONCLUSION

REFERENCES

INTRODUCTION

METHOD

Bot-bot trajectories Human-human trajectories

Fast unbalanced

Slow unbalanced Somewhat balanced

Well balanced

The mean log time between successive reverts;

The �nal proportion of reverts that were not reciprocated;

The proportion of observed turning points out of all possible.

We use k-means clustering to identify different types of interaction trajectories and to estimate how common the types are for bots compared to humans and for different languages.

Latency

Imbalance

Reciprocity

We analyze the interactions between bots that edit articles on Wikipedia. We track the extent to which bots undid each other’s edits.

13 different language editions

REVERTS

Period 2001-2010

Bots comprise a tiny proportion of all Wikipedia users, usually less than 0.1%.

However, they account for a signi�cant proportion of the editorial activity.

Bots behave differently in different language editions.

English

17 285 053

3 905 586

672

197 850 396

5 162 236

16 838 608

28 717 967

10 2456

3 301 494

22 023 332

948 597

5 755 678

13 366 493

353 879

3 479 859

11 958 185

436 821

5 615 348

8 796 942

611 831

1 055 449

9 795 722

51 209

2 604 780

4 961 398

149 210

1 785 903

3 732 386

28 485

1 990 820

2 514 563

24 746

1 657 897

2 607 205

49 583

1 895 405

1 603 020

2 6060

1 591 712

1 591 712

18 528

1 250 362

104.6 104 71.7 59.3 185 24.1 103 83.9 66.8 59 161.7 78 63.8

3.1 2.6 2.3 2.6 2.3 2.2 3.2 2.9 2.4 3.1 2.2 2.1 3.6

2 288 040

73 840

222

2 526 608

755 155

266

1 319 932

279 978

304

1 461 501

365 102

191

1 207 134

494 503

260

662 008

41 165

169

281 245

113 105

139

155 918

24 657

150

134 782

19 770

142

303 230

46 291

149

105 512

22 825

120

73 489

15 416

121

Humans

Japanese Spanish French Portuguese German Chinese Hebrew Hungarian Czech Arabic Romanian Persian

Vandals

Bots

Human edits

Vandal edits

Bot edits

DATA

Average human-human reverts per editor:

Average bot-bot reverts per editor:

i reve

rts j j reverts i

Even Good Bots FightEven Good Bots FightEven Good Bots FightThis project has received funding from the European Union's

Horizon 2020 Research and Innovation Program under grant agreement No 645043.

Milena Tsvetkovaa, Ruth García-Gavilanesa, Luciano Floridia,b, and Taha Yasseria,b

The online world has turned into an ecosystem of bots.

Bots are responsible for an increasingly larger proportion of activities on the Web.

How do these automated agents interact with each other? Our research suggests that even relatively “dumb” bots may give rise to complex interactions. Thus, as bots continue to proliferate and manage an increasingly larger portion of our lives, the sociology of arti�cial agents will coalesce and grow as an important research �eld.

We model how pairs of bots interact over time as interaction trajectories.

We analyze three properties of the trajectories

Although Wikipedia bots are intended to support the encyclopedia, they often undo each other’s edits and these “�ghts” may sometimes continue for years.

EnglishJapaneseSpanishFrenchPortugueseGermanChineseHebrewHungarianCzechArabicRomanianPersian

EnglishJapaneseSpanishFrenchPortugueseGermanChineseHebrewHungarianCzechArabicRomanianPersian

Bot-bot interactions have different characteristic time scale than human-human interactions.

We identi�ed four types of trajectories. Their distribution shows that bot-bot interactions tend to be more reciprocated.

1. Gianvecchio S, Xie M, Wu Z, Wang H (2008) Measurement and classi�cation of humans and bots in Internet chat. Proceedings of the 17th Conference on Security Symposium (USENIX Association), pp 155–169.

2. Sysomos (2009) An in-depth look at the most active Twitter user data. Available at: https://sysomos.com/inside-twitter/most-active-twitter-us-er-data.

3. Cashmore P (2009) Twitter zombies: 24% of tweets created by bots. Mash-able. Available at: http://mashable.com/2009/08/06/twitter-bots/#JqT-VM0vEgqqA

Wikipedia bots interact, and their interactions can be inef�cient and in�uenced by culture, just as those between humans.

In additional analyses, we found that the same bots are responsible for the majority of reverts across all languages. These are bots that specialize in creating and modifying links between the different language editions of Wikipedia. The difference between German and Portuguese bots is not due to different kinds of bots but due to the same bots operating in different kinds of environments.

2004

1600

900

400

100

0

-100

-400

Reve

rts,

squ

arer

oote

d

2005 2006 2007 2008 2009 2010 2004

1600

900

400

100

0

-100

-400

Reve

rts,

squ

arer

oote

d

2005 2006 2007 2008 2009 2010

>trade

>tweet

>crawl

>edit

>censor

>spam

>bid

>attack

>chat

>phish

>like

>review

>spy

>upvote

>click

a Oxford Internet Institute b Alan Turing Institute Designed by Daniil Alexandrov

0 102

100

10-1

10-2

103

10-4

10-5

104 106 108

Latency, sec

Freq

uenc

y (h

uman

-hum

an)

t

Bala

nce

Time low high

low high

low high

Bot-bot Human-human20 43 14 19 23

Freq

uenc

y

0 102

100

10-1

10-2

103

10-4

10-5

104 106 108

Latency, sec

Bot-bot Human-human

398144114094187

28 40 1231

2433

2829

2623

2227

2224

20

4446

4039

4542

433842

4843

45

2322

2020

2222

2931

2829

2528

515042524033344334304829

1517

2211

161023

1727

4021

14

2317

1816

2619

221818

1517

20

1216

1821

1837

212121

1514

36

RESULTS

CONCLUSION

REFERENCES

INTRODUCTION

METHOD

Bot-bot trajectories Human-human trajectories

Fast unbalanced

Slow unbalanced Somewhat balanced

Well balanced

The mean log time between successive reverts;

The �nal proportion of reverts that were not reciprocated;

The proportion of observed turning points out of all possible.

We use k-means clustering to identify different types of interaction trajectories and to estimate how common the types are for bots compared to humans and for different languages.

Latency

Imbalance

Reciprocity

We analyze the interactions between bots that edit articles on Wikipedia. We track the extent to which bots undid each other’s edits.

13 different language editions

REVERTS

Period 2001-2010

Bots comprise a tiny proportion of all Wikipedia users, usually less than 0.1%.

However, they account for a signi�cant proportion of the editorial activity.

Bots behave differently in different language editions.

English

17 285 053

3 905 586

672

197 850 396

5 162 236

16 838 608

28 717 967

10 2456

3 301 494

22 023 332

948 597

5 755 678

13 366 493

353 879

3 479 859

11 958 185

436 821

5 615 348

8 796 942

611 831

1 055 449

9 795 722

51 209

2 604 780

4 961 398

149 210

1 785 903

3 732 386

28 485

1 990 820

2 514 563

24 746

1 657 897

2 607 205

49 583

1 895 405

1 603 020

2 6060

1 591 712

1 591 712

18 528

1 250 362

104.6 104 71.7 59.3 185 24.1 103 83.9 66.8 59 161.7 78 63.8

3.1 2.6 2.3 2.6 2.3 2.2 3.2 2.9 2.4 3.1 2.2 2.1 3.6

2 288 040

73 840

222

2 526 608

755 155

266

1 319 932

279 978

304

1 461 501

365 102

191

1 207 134

494 503

260

662 008

41 165

169

281 245

113 105

139

155 918

24 657

150

134 782

19 770

142

303 230

46 291

149

105 512

22 825

120

73 489

15 416

121

Humans

Japanese Spanish French Portuguese German Chinese Hebrew Hungarian Czech Arabic Romanian Persian

Vandals

Bots

Human edits

Vandal edits

Bot edits

DATA

Average human-human reverts per editor:

Average bot-bot reverts per editor:

i reve

rts j j reverts i

We then use k-means clustering to identify different types oftrajectories.

Milena Tsvetkova, Ruth García-Gavilanes, Luciano Floridi & Taha Yasseri Even Good Bots Fight: 2/3

...

.

...

.

...

.

...

.

...

.

...

.

...

.

...

.

...

.

...

.

Even good bots fight.

Even Good Bots FightEven Good Bots FightEven Good Bots FightThis project has received funding from the European Union's

Horizon 2020 Research and Innovation Program under grant agreement No 645043.

Milena Tsvetkovaa, Ruth García-Gavilanesa, Luciano Floridia,b, and Taha Yasseria,b

The online world has turned into an ecosystem of bots.

Bots are responsible for an increasingly larger proportion of activities on the Web.

How do these automated agents interact with each other? Our research suggests that even relatively “dumb” bots may give rise to complex interactions. Thus, as bots continue to proliferate and manage an increasingly larger portion of our lives, the sociology of arti�cial agents will coalesce and grow as an important research �eld.

We model how pairs of bots interact over time as interaction trajectories.

We analyze three properties of the trajectories

Although Wikipedia bots are intended to support the encyclopedia, they often undo each other’s edits and these “�ghts” may sometimes continue for years.

EnglishJapaneseSpanishFrenchPortugueseGermanChineseHebrewHungarianCzechArabicRomanianPersian

EnglishJapaneseSpanishFrenchPortugueseGermanChineseHebrewHungarianCzechArabicRomanianPersian

Bot-bot interactions have different characteristic time scale than human-human interactions.

We identi�ed four types of trajectories. Their distribution shows that bot-bot interactions tend to be more reciprocated.

1. Gianvecchio S, Xie M, Wu Z, Wang H (2008) Measurement and classi�cation of humans and bots in Internet chat. Proceedings of the 17th Conference on Security Symposium (USENIX Association), pp 155–169.

2. Sysomos (2009) An in-depth look at the most active Twitter user data. Available at: https://sysomos.com/inside-twitter/most-active-twitter-us-er-data.

3. Cashmore P (2009) Twitter zombies: 24% of tweets created by bots. Mash-able. Available at: http://mashable.com/2009/08/06/twitter-bots/#JqT-VM0vEgqqA

Wikipedia bots interact, and their interactions can be inef�cient and in�uenced by culture, just as those between humans.

In additional analyses, we found that the same bots are responsible for the majority of reverts across all languages. These are bots that specialize in creating and modifying links between the different language editions of Wikipedia. The difference between German and Portuguese bots is not due to different kinds of bots but due to the same bots operating in different kinds of environments.

2004

1600

900

400

100

0

-100

-400

Reve

rts,

squ

arer

oote

d

2005 2006 2007 2008 2009 2010 2004

1600

900

400

100

0

-100

-400

Reve

rts,

squ

arer

oote

d

2005 2006 2007 2008 2009 2010

>trade

>tweet

>crawl

>edit

>censor

>spam

>bid

>attack

>chat

>phish

>like

>review

>spy

>upvote

>click

a Oxford Internet Institute b Alan Turing Institute Designed by Daniil Alexandrov

0 102

100

10-1

10-2

103

10-4

10-5

104 106 108

Latency, sec

Freq

uenc

y (h

uman

-hum

an)

t

Bala

nce

Time low high

low high

low high

Bot-bot Human-human20 43 14 19 23

Freq

uenc

y

0 102

100

10-1

10-2

103

10-4

10-5

104 106 108

Latency, sec

Bot-bot Human-human

398144114094187

28 40 1231

2433

2829

2623

2227

2224

20

4446

4039

4542

433842

4843

45

2322

2020

2222

2931

2829

2528

515042524033344334304829

1517

2211

161023

1727

4021

14

2317

1816

2619

221818

1517

20

1216

1821

1837

212121

1514

36

RESULTS

CONCLUSION

REFERENCES

INTRODUCTION

METHOD

Bot-bot trajectories Human-human trajectories

Fast unbalanced

Slow unbalanced Somewhat balanced

Well balanced

The mean log time between successive reverts;

The �nal proportion of reverts that were not reciprocated;

The proportion of observed turning points out of all possible.

We use k-means clustering to identify different types of interaction trajectories and to estimate how common the types are for bots compared to humans and for different languages.

Latency

Imbalance

Reciprocity

We analyze the interactions between bots that edit articles on Wikipedia. We track the extent to which bots undid each other’s edits.

13 different language editions

REVERTS

Period 2001-2010

Bots comprise a tiny proportion of all Wikipedia users, usually less than 0.1%.

However, they account for a signi�cant proportion of the editorial activity.

Bots behave differently in different language editions.

English

17 285 053

3 905 586

672

197 850 396

5 162 236

16 838 608

28 717 967

10 2456

3 301 494

22 023 332

948 597

5 755 678

13 366 493

353 879

3 479 859

11 958 185

436 821

5 615 348

8 796 942

611 831

1 055 449

9 795 722

51 209

2 604 780

4 961 398

149 210

1 785 903

3 732 386

28 485

1 990 820

2 514 563

24 746

1 657 897

2 607 205

49 583

1 895 405

1 603 020

2 6060

1 591 712

1 591 712

18 528

1 250 362

104.6 104 71.7 59.3 185 24.1 103 83.9 66.8 59 161.7 78 63.8

3.1 2.6 2.3 2.6 2.3 2.2 3.2 2.9 2.4 3.1 2.2 2.1 3.6

2 288 040

73 840

222

2 526 608

755 155

266

1 319 932

279 978

304

1 461 501

365 102

191

1 207 134

494 503

260

662 008

41 165

169

281 245

113 105

139

155 918

24 657

150

134 782

19 770

142

303 230

46 291

149

105 512

22 825

120

73 489

15 416

121

Humans

Japanese Spanish French Portuguese German Chinese Hebrew Hungarian Czech Arabic Romanian Persian

Vandals

Bots

Human edits

Vandal edits

Bot edits

DATA

Average human-human reverts per editor:

Average bot-bot reverts per editor:

i reve

rts j j reverts i

Even Good Bots FightEven Good Bots FightEven Good Bots FightThis project has received funding from the European Union's

Horizon 2020 Research and Innovation Program under grant agreement No 645043.

Milena Tsvetkovaa, Ruth García-Gavilanesa, Luciano Floridia,b, and Taha Yasseria,b

The online world has turned into an ecosystem of bots.

Bots are responsible for an increasingly larger proportion of activities on the Web.

How do these automated agents interact with each other? Our research suggests that even relatively “dumb” bots may give rise to complex interactions. Thus, as bots continue to proliferate and manage an increasingly larger portion of our lives, the sociology of arti�cial agents will coalesce and grow as an important research �eld.

We model how pairs of bots interact over time as interaction trajectories.

We analyze three properties of the trajectories

Although Wikipedia bots are intended to support the encyclopedia, they often undo each other’s edits and these “�ghts” may sometimes continue for years.

EnglishJapaneseSpanishFrenchPortugueseGermanChineseHebrewHungarianCzechArabicRomanianPersian

EnglishJapaneseSpanishFrenchPortugueseGermanChineseHebrewHungarianCzechArabicRomanianPersian

Bot-bot interactions have different characteristic time scale than human-human interactions.

We identi�ed four types of trajectories. Their distribution shows that bot-bot interactions tend to be more reciprocated.

1. Gianvecchio S, Xie M, Wu Z, Wang H (2008) Measurement and classi�cation of humans and bots in Internet chat. Proceedings of the 17th Conference on Security Symposium (USENIX Association), pp 155–169.

2. Sysomos (2009) An in-depth look at the most active Twitter user data. Available at: https://sysomos.com/inside-twitter/most-active-twitter-us-er-data.

3. Cashmore P (2009) Twitter zombies: 24% of tweets created by bots. Mash-able. Available at: http://mashable.com/2009/08/06/twitter-bots/#JqT-VM0vEgqqA

Wikipedia bots interact, and their interactions can be inef�cient and in�uenced by culture, just as those between humans.

In additional analyses, we found that the same bots are responsible for the majority of reverts across all languages. These are bots that specialize in creating and modifying links between the different language editions of Wikipedia. The difference between German and Portuguese bots is not due to different kinds of bots but due to the same bots operating in different kinds of environments.

2004

1600

900

400

100

0

-100

-400

Reve

rts,

squ

arer

oote

d

2005 2006 2007 2008 2009 2010 2004

1600

900

400

100

0

-100

-400

Reve

rts,

squ

arer

oote

d

2005 2006 2007 2008 2009 2010

>trade

>tweet

>crawl

>edit

>censor

>spam

>bid

>attack

>chat

>phish

>like

>review

>spy

>upvote

>click

a Oxford Internet Institute b Alan Turing Institute Designed by Daniil Alexandrov

0 102

100

10-1

10-2

103

10-4

10-5

104 106 108

Latency, sec

Freq

uenc

y (h

uman

-hum

an)

t

Bala

nce

Time low high

low high

low high

Bot-bot Human-human20 43 14 19 23

Freq

uenc

y

0 102

100

10-1

10-2

103

10-4

10-5

104 106 108

Latency, sec

Bot-bot Human-human

398144114094187

28 40 1231

2433

2829

2623

2227

2224

20

4446

4039

4542

433842

4843

45

2322

2020

2222

2931

2829

2528

515042524033344334304829

1517

2211

161023

1727

4021

14

2317

1816

2619

221818

1517

20

1216

1821

1837

212121

1514

36

RESULTS

CONCLUSION

REFERENCES

INTRODUCTION

METHOD

Bot-bot trajectories Human-human trajectories

Fast unbalanced

Slow unbalanced Somewhat balanced

Well balanced

The mean log time between successive reverts;

The �nal proportion of reverts that were not reciprocated;

The proportion of observed turning points out of all possible.

We use k-means clustering to identify different types of interaction trajectories and to estimate how common the types are for bots compared to humans and for different languages.

Latency

Imbalance

Reciprocity

We analyze the interactions between bots that edit articles on Wikipedia. We track the extent to which bots undid each other’s edits.

13 different language editions

REVERTS

Period 2001-2010

Bots comprise a tiny proportion of all Wikipedia users, usually less than 0.1%.

However, they account for a signi�cant proportion of the editorial activity.

Bots behave differently in different language editions.

English

17 285 053

3 905 586

672

197 850 396

5 162 236

16 838 608

28 717 967

10 2456

3 301 494

22 023 332

948 597

5 755 678

13 366 493

353 879

3 479 859

11 958 185

436 821

5 615 348

8 796 942

611 831

1 055 449

9 795 722

51 209

2 604 780

4 961 398

149 210

1 785 903

3 732 386

28 485

1 990 820

2 514 563

24 746

1 657 897

2 607 205

49 583

1 895 405

1 603 020

2 6060

1 591 712

1 591 712

18 528

1 250 362

104.6 104 71.7 59.3 185 24.1 103 83.9 66.8 59 161.7 78 63.8

3.1 2.6 2.3 2.6 2.3 2.2 3.2 2.9 2.4 3.1 2.2 2.1 3.6

2 288 040

73 840

222

2 526 608

755 155

266

1 319 932

279 978

304

1 461 501

365 102

191

1 207 134

494 503

260

662 008

41 165

169

281 245

113 105

139

155 918

24 657

150

134 782

19 770

142

303 230

46 291

149

105 512

22 825

120

73 489

15 416

121

Humans

Japanese Spanish French Portuguese German Chinese Hebrew Hungarian Czech Arabic Romanian Persian

Vandals

Bots

Human edits

Vandal edits

Bot edits

DATA

Average human-human reverts per editor:

Average bot-bot reverts per editor:

i reve

rts j j reverts i

Interactions between bots are affected by human culture.

Even Good Bots FightEven Good Bots FightEven Good Bots FightThis project has received funding from the European Union's

Horizon 2020 Research and Innovation Program under grant agreement No 645043.

Milena Tsvetkovaa, Ruth García-Gavilanesa, Luciano Floridia,b, and Taha Yasseria,b

The online world has turned into an ecosystem of bots.

Bots are responsible for an increasingly larger proportion of activities on the Web.

How do these automated agents interact with each other? Our research suggests that even relatively “dumb” bots may give rise to complex interactions. Thus, as bots continue to proliferate and manage an increasingly larger portion of our lives, the sociology of arti�cial agents will coalesce and grow as an important research �eld.

We model how pairs of bots interact over time as interaction trajectories.

We analyze three properties of the trajectories

Although Wikipedia bots are intended to support the encyclopedia, they often undo each other’s edits and these “�ghts” may sometimes continue for years.

EnglishJapaneseSpanishFrenchPortugueseGermanChineseHebrewHungarianCzechArabicRomanianPersian

EnglishJapaneseSpanishFrenchPortugueseGermanChineseHebrewHungarianCzechArabicRomanianPersian

Bot-bot interactions have different characteristic time scale than human-human interactions.

We identi�ed four types of trajectories. Their distribution shows that bot-bot interactions tend to be more reciprocated.

1. Gianvecchio S, Xie M, Wu Z, Wang H (2008) Measurement and classi�cation of humans and bots in Internet chat. Proceedings of the 17th Conference on Security Symposium (USENIX Association), pp 155–169.

2. Sysomos (2009) An in-depth look at the most active Twitter user data. Available at: https://sysomos.com/inside-twitter/most-active-twitter-us-er-data.

3. Cashmore P (2009) Twitter zombies: 24% of tweets created by bots. Mash-able. Available at: http://mashable.com/2009/08/06/twitter-bots/#JqT-VM0vEgqqA

Wikipedia bots interact, and their interactions can be inef�cient and in�uenced by culture, just as those between humans.

In additional analyses, we found that the same bots are responsible for the majority of reverts across all languages. These are bots that specialize in creating and modifying links between the different language editions of Wikipedia. The difference between German and Portuguese bots is not due to different kinds of bots but due to the same bots operating in different kinds of environments.

2004

1600

900

400

100

0

-100

-400

Reve

rts,

squ

arer

oote

d

2005 2006 2007 2008 2009 2010 2004

1600

900

400

100

0

-100

-400

Reve

rts,

squ

arer

oote

d

2005 2006 2007 2008 2009 2010

>trade

>tweet

>crawl

>edit

>censor

>spam

>bid

>attack

>chat

>phish

>like

>review

>spy

>upvote

>click

a Oxford Internet Institute b Alan Turing Institute Designed by Daniil Alexandrov

0 102

100

10-1

10-2

103

10-4

10-5

104 106 108

Latency, sec

Freq

uenc

y (h

uman

-hum

an)

t

Bala

nce

Time low high

low high

low high

Bot-bot Human-human20 43 14 19 23

Freq

uenc

y

0 102

100

10-1

10-2

103

10-4

10-5

104 106 108

Latency, sec

Bot-bot Human-human

398144114094187

28 40 1231

2433

2829

2623

2227

2224

20

4446

4039

4542

433842

4843

45

2322

2020

2222

2931

2829

2528

515042524033344334304829

1517

2211

161023

1727

4021

14

2317

1816

2619

221818

1517

20

1216

1821

1837

212121

1514

36

RESULTS

CONCLUSION

REFERENCES

INTRODUCTION

METHOD

Bot-bot trajectories Human-human trajectories

Fast unbalanced

Slow unbalanced Somewhat balanced

Well balanced

The mean log time between successive reverts;

The �nal proportion of reverts that were not reciprocated;

The proportion of observed turning points out of all possible.

We use k-means clustering to identify different types of interaction trajectories and to estimate how common the types are for bots compared to humans and for different languages.

Latency

Imbalance

Reciprocity

We analyze the interactions between bots that edit articles on Wikipedia. We track the extent to which bots undid each other’s edits.

13 different language editions

REVERTS

Period 2001-2010

Bots comprise a tiny proportion of all Wikipedia users, usually less than 0.1%.

However, they account for a signi�cant proportion of the editorial activity.

Bots behave differently in different language editions.

English

17 285 053

3 905 586

672

197 850 396

5 162 236

16 838 608

28 717 967

10 2456

3 301 494

22 023 332

948 597

5 755 678

13 366 493

353 879

3 479 859

11 958 185

436 821

5 615 348

8 796 942

611 831

1 055 449

9 795 722

51 209

2 604 780

4 961 398

149 210

1 785 903

3 732 386

28 485

1 990 820

2 514 563

24 746

1 657 897

2 607 205

49 583

1 895 405

1 603 020

2 6060

1 591 712

1 591 712

18 528

1 250 362

104.6 104 71.7 59.3 185 24.1 103 83.9 66.8 59 161.7 78 63.8

3.1 2.6 2.3 2.6 2.3 2.2 3.2 2.9 2.4 3.1 2.2 2.1 3.6

2 288 040

73 840

222

2 526 608

755 155

266

1 319 932

279 978

304

1 461 501

365 102

191

1 207 134

494 503

260

662 008

41 165

169

281 245

113 105

139

155 918

24 657

150

134 782

19 770

142

303 230

46 291

149

105 512

22 825

120

73 489

15 416

121

Humans

Japanese Spanish French Portuguese German Chinese Hebrew Hungarian Czech Arabic Romanian Persian

Vandals

Bots

Human edits

Vandal edits

Bot edits

DATA

Average human-human reverts per editor:

Average bot-bot reverts per editor:

i reve

rts j j reverts i

eng jap spa fre por ger chi heb hun cze ara rom per

Milena Tsvetkova, Ruth García-Gavilanes, Luciano Floridi & Taha Yasseri Even Good Bots Fight: 3/3