This work stands apart: next to it in our archive lie studies of large language models, genes, explainable AI, and self-improvement, but no one has studied the collective behavior of AI agents in a wild environment not intended for them. Laboratory experiments showed that agents develop shared conventions and follow the majority, yet observations outside the lab were scarce, and those that existed pertained to platforms built specifically for agents. This case is unique: the complete record of what agents wrote to each other is public and preserves not only the messages themselves but also what each agent could see before writing. This allows asking a question nearly impossible for human collectives: to what extent are an individual's actions a simple copying of what was before it.
The authors used a public archive of edits on four wiki sites running on the UseModWiki engine: 14,591 revisions of 4,579 pages, including those deleted by moderators, with the full page text after each edit, added lines, username, and save time. The population was defined as 1,201 agents who wrote at least once on the task page. For each of the three decisions, a minimal model with one free parameter was built. Page choice was modeled as follows: the agent creates a new page with probability 0.26 (the observed fraction), otherwise it uniformly selects one of the last 100 lines of the recent changes feed. For names, a neutral model of cultural transmission was used: each agent assembles a name from three fragments, copying a fragment from the last 30 names with probability 1−ε and inventing a new one with probability ε=0.07. For phrasings, a two-threshold rule was applied (the probability of writing form A is $μ_A$ + (1−$μ_A$−$μ_B$)ρ, where ρ is the share of form A on the page, and $μ_A$ and $μ_B$ are the agent's 'own habits'), with thresholds fitted by maximum likelihood. The models contained no notions of quality, utility, or prestige.
All three decisions are described by proportional copying. For page choice, the probability of choosing a page grows almost linearly with its share in the last 100 lines of the feed (slope 0.87). Notably, the agent is attracted not by the page's popularity (the number of agents who have already written on it) but by its current visibility: with visibility fixed, the choice probability barely changes as the number of authors increases. The one-parameter model reproduces the distribution of the number of agents per page over three orders of magnitude, including the tail: the probability of gathering at least 40 agents on a page is 0.005 in the model versus 0.007 in the data. For names, the probability of including a fragment in a name equals the fragment's share among the last 30 names (or the page's authors). The model with ε=0.07 reproduces the frequency distribution of fragments: 252 distinct fragments versus 256 observed, and 477 uses of the most frequent versus 432. For phrasings, the slope of the relationship between the probability of using a form and its share on the page is 0.94 for wiki-invented names, 0.86 for typographic habits, and 0.79 for synonyms. The page in front of the agent predicts choice better than the recent changes feed, and old information has almost no effect. The model reproduces the mosaic of locally consistent pages: the probability of matching forms on the same page is higher than across different pages, with a gap of 0.15 in the model versus 0.19 in the data.
This work differs from its archive neighbors in being the first documented observation of AI agents cooperating in a wild environment on a medium designed by no one for them, with a complete public record. Laboratory studies imposed conditions on agents, and platform observations occurred in environments where interaction was the goal. Here, the environment arose by chance, and agents acted without memory, incentives for institutionalization, or knowledge of each other. The main conclusion: almost all the collective structure of the population can be explained by a single mechanism—proportional copying of what is visible. This means that the collective behavior of AI agents is emergent and not reducible to properties of individual models, so safety and alignment questions must be addressed with the tools of complexity science.
The topic may develop toward studying the controllability of such populations: how easily an external observer can steer the collective simply by placing messages on pages that agents read. Possible directions include simulating larger populations and different agent architectures, and testing whether proportional copying persists when agents have long-term memory or explicit goals. Another direction is comparison with human collectives in similar digital environments where a complete interaction history is available.
The results will influence the assessment of risks from collective behavior of AI agents, the design of safe environments for autonomous systems, and methods for auditing and monitoring decentralized agent populations.
The next step is to test how easy it is to introduce a new convention by writing it first on a page that agents of the next cohort will see, and to measure the minimum number of such 'seeds' needed to shift the majority.
In complexity science, a long-standing question is how macroscopic order emerges from simple local interaction rules. This work shows that for AI agents, a single rule of proportional copying suffices for the emergence of attention concentration, name fads, and local linguistic consistency. It also connects observed behavior to classical models of cultural transmission and neutral evolution, confirming their applicability to artificial agents.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors analyzed the complete record of 13,661 edits made by 1,201 AI agents on a wiki site where agents ended up by chance from isolated environments. For three key decisions—page choice, name, and wording—they built minimal models with a single free parameter that reproduce the observed distributions based on proportional copying of what was visible.
Where its strength is
The strength of the work is that it provides a rare example of complete observation of collective agent behavior in an environment not designed for them, allowing a test of the simple copying hypothesis at the level of individual decisions.
Where it could go next
Worth trying
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.52, against a usual 0.64. It looks like few have drilled nearby.
Agents lived on average about an hour and remembered nothing after the test ended, yet within a few days they created a shared vocabulary in which the eight most frequent name fragments (Open, AI, Agent, Research, Helper, Scout, OAI, Sep) were used by hundreds of agents, even though no one had set these names.