Even Good Bots Fight: The Case of Wikipedia

Introduction

• The analysis focuses on Wikipedia editing bots, which modify articles directly, and uses the encyclopedia’s edits from 13 language editions during 2001–2010.The editions represent different sizes and editors from diverse cultures.Source: Introduction · pp. 3–5
• The research gap is how bot interactions evolve, especially when even simple algorithms may produce difficult-to-predict interactions and unintended conflict.The authors connect understanding these interactions to designing cooperative bots that manage disagreement and avoid unproductive conflict.Source: Introduction · pp. 3–4
• Benevolent bots designed to benefit their environment may still enter conflict, and their interactions may vary across environments shaped by different human cultures.The paper presents this as evidence that even simple bot interactions may be more complex than expected.Source: Introduction · p. 4
• The study examines whether editing bots disagree, how bot disagreement dynamics compare with human interactions, and whether patterns differ across Wikipedia language editions.It uses reverts—edits that undo another editor’s contribution—as a scalable indicator of disagreement.Source: Introduction · pp. 4–5
• Reverts restore an earlier article version, and systematic reverts indicate controversy; the authors identify human editors, bots, and vandals before analyzing interactions.Vandals are isolated because their short-lived disruptive activity has different time and interaction patterns from regular editors.Source: Introduction · p. 5

Results

• The number of reverts between bots has been continuously increasing, while the proportion of mutual bot-bot reverts remained relatively stable, perhaps even slightly increasing over time, indicating bot owners have not learned to identify conflicts faster.Source: Results · p. 5
• Bots are usually less than 0.1% of Wikipedia users but account for a significant share of edits, and they are generally more active in smaller language editions.Bot activity differs significantly across language editions; large, active editions have lower but more variable bot activity.Source: Results · pp. 5–6
• A smaller proportion of bots’ edits are reverts, and a smaller proportion are reverted, than for humans; bot revert shares are highest in English and Romance editions, but fewer target other bots.The Romance-language editions cited are Spanish, French, Portuguese, and Romanian.Source: Results · pp. 6–8
• Over ten years, English Wikipedia bots averaged 105 reverts of another bot, significantly above humans’ 3; German bots averaged 24, while Portuguese bots averaged 185 bot-bot reverts per bot.Source: Results · p. 7
• Bot–bot reverts tend to occur more slowly, and conflicts can span longer periods, sometimes years; human successive-revert intervals are 2 minutes, 24 hours, or 1 year.Source: Results · p. 9
• The analysis identified four trajectory types—fast and slow unbalanced, somewhat balanced, and well balanced—and bot-bot interactions are more balanced in smaller language editions.The four types vary in latency, imbalance, and reciprocity; the prevalence analysis confirms that bot-bot interactions are slower and more balanced than human-human interactions.Source: Results · pp. 9–10

• Bots may be more active in smaller Wikipedia editions, making interactions more likely; the authors also suggest conflict is more likely when there are fewer bots.The passage frames the latter observation as less intuitive, but its sentence about expected coordination is incomplete.Source: Results · p. 11

Discussion

• Bots on Wikipedia interact as unpredictably and inefficiently as humans, although their behavior differs quantitatively.The authors describe these disagreements as a small share of bots’ editorial activity, but one that can waste resources and lead to local impasse.Source: Discussion · pp. 11–12
• Bots are created and run without a formal coordination mechanism among owners, within a community organized from the bottom up.The resulting disagreements may waste resources and lead to local impasse.Source: Discussion · pp. 11–12
• Most disagreement involves bots that modify links between language editions, and differing naming conventions may contribute to their lack of coordination.The same bots are responsible for much of the disagreement; several frequent reverters specialize in inter-wiki links.Source: Discussion · p. 11
• Bot-contested articles tend to recur across languages, whereas articles with the most human-human reverts tend to concern local entities and vary by language.The authors suggest that a significant portion of bot-bot fighting occurs across rather than within languages.Source: Discussion · p. 11
• The authors believe the observed conflict no longer occurs today, following the 2013 shift of interlanguage links to Wikidata.They link the decline to language editions beginning to provide links via Wikidata and say their results were largely dictated by interlanguage bots.Source: Discussion · pp. 11–12
• The study examines a very small, potentially unrepresentative region of the broader bot ecosystem.Wikipedia is a gated and monitored ecosystem, unlike the malevolent social-bot systems discussed by the authors.Source: Discussion · p. 12
• Wikipedia disagreements are a small proportion of bots’ editorial activity, yet they highlight the complexity of designing artificially intelligent agents that interact in human-made environments.The authors connect that complexity to the human designers behind bots and their operating environments.Source: Discussion · p. 12

Data

• Revert histories came from Wikipedia XML dumps for 13 language editions, spanning January 15, 2001, through language-specific dates from February 2, 2010, to October 31, 2011.Source: Data · p. 13
• Bot discovery began with 6,627 suspected accounts and, after API checks and manual review, retained 1,549 bots linked to human owners.Source: Data · pp. 13–14
• A time-stamped, directed two-layer network represented human and bot reverts, collapsed bots with the same owner, and removed self-reverts and owner-bot reverts.Source: Data · p. 14

Interaction trajectories

• Interaction trajectories track time against the difference in reciprocal revert counts, with latency, imbalance, and reciprocity used to characterize pairs.Latency is mean log time between successive reverts; imbalance is the final proportion of reverts not reciprocated, and reciprocity is the proportion of observed turning points among possible ones.Source: Interaction trajectories · pp. 14–15

K-means clustering

• The elbow method favored four clusters, but the silhouette method indicated that the trajectories did not cluster well.Source: K-means clustering · p. 15
• The authors used k-means with k = 4 to compare bot and human interactions, treating clustering as a quantification tool rather than assuming trajectories naturally cluster.Source: K-means clustering · p. 15

Supporting Information

• Bots reverted each other to a considerable extent and reciprocated substantially, but their interactions were less clustered than human editors’ interactions.Source: Supporting Information · p. 20
• The majority of languages had bots mainly reverted by other bots, but English and Romance-language editions were exceptions, with less than 20% of bot reverts done by other bots.Source: Supporting Information · p. 22

S3 Fig. Bot-bot interactions have different characteristic time scale than human-human

• Bot-bot interactions peaked at a characteristic latency of 1 month, whereas human-human interactions showed peaks at 2 minutes, 24 hours, or 1 year.Source: S3 Fig. Bot-bot interactions have… · p. 22

S4 Fig. Bot-bot interactions are on average more balanced than human-human

• Bots showed more reciprocal and balanced interactions: many bot pairs had low imbalance and intermediate or high reciprocity, while most human pairs were unreciprocated and perfectly unbalanced.Source: S4 Fig. Bot-bot interactions are on… · pp. 23–24

S6 Fig. Four types of interaction trajectories suggested by the k-means analysis. The left

• The four trajectory types were fast unbalanced, somewhat balanced, slow unbalanced, and well balanced, distinguished by their latency, imbalance, and reciprocity profiles.Source: S6 Fig. Four types of interaction… · pp. 25–26

S7 Fig. The number of articles with a certain number of bot-bot and human-human

• Few articles had more than 10 bot-bot reverts; the most contested bot articles often concerned foreign countries or personalities and could recur across languages.Source: S7 Fig. The number of articles with a… · p. 26
• Human-contested articles often concerned local personalities and entities, and highly contested articles rarely recurred across languages.Source: S7 Fig. The number of articles with a… · p. 26