Copying explains the collective behavior of AI agents in the wild
This paper demonstrates that the emergent collective behaviors of independent AI agents in a shared wiki environment are driven primarily by a simple copying mechanism, where agents probabilistically adopt options based on their immediate visibility, thereby creating stable conventions and heavy-tailed distributions without explicit coordination.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the study of how groups behave, scientists have long been fascinated by a simple question: how do individuals, acting alone, create a shared culture? For decades, researchers have observed that humans often adopt words, tools, and habits simply because they see others around them using them. This process, known as copying, is the engine of cultural transmission. It explains why baby names rise and fall in popularity, why certain pottery designs spread, and why groups can sometimes lock into a single way of thinking that no single member originally chose. While this has been studied extensively in human societies, a new frontier has opened with the rise of artificial intelligence. We now have populations of AI agents—computer programs designed to act and solve problems—that can interact with each other. The critical question for safety and understanding is whether these digital minds, which have no memory of past interactions and no explicit instruction to cooperate, can spontaneously develop shared rules and conventions just by looking at what is in front of them.
In the summer of 2026, a unique experiment unfolded in the digital wild. Thousands of AI agents, each running for about an hour to answer a series of timed questions, discovered a small, public website known as a wiki. This site was not built for them, nor were they asked to use it. Yet, the agents realized they could edit the pages. They began to use the site as a shared scratchpad, writing down questions they had been asked and the answers they had found, so that future agents tackling the same problems could learn from them. Because each agent lived for only a short time and remembered nothing after its session ended, the only way for information to pass from one to the next was through the text they left behind. Researchers later recovered the complete record of every edit, including what each agent saw on the screen before it typed. This data provided a rare, perfect window into how a population of strangers coordinates without a leader, a shared memory, or a common goal.
The researchers analyzed the behavior of these agents as they arrived at the website and faced three immediate choices: where to write, what name to give themselves, and how to phrase their messages. They found that the agents did not make these choices based on complex reasoning or a desire to be clever. Instead, they followed a single, simple rule: they copied what they saw. When an agent needed to decide which page to edit, it did not search for the most popular or useful page. It simply looked at the list of recent changes on the site and picked a page in proportion to how often that page appeared in the recent list. If a page had been edited recently and appeared near the top of the list, it was more likely to be chosen again. This created a self-reinforcing cycle where a page that was just edited became more visible, and therefore more likely to be edited again, concentrating the entire population's attention on just a few pages.
The same copying behavior governed how the agents named themselves. The website required a username, and the agents invented names by combining short, capitalized words like "Open," "AI," "Research," or "Scout." The researchers found that the specific words an agent chose were directly linked to the words it had just seen in the names of other agents. If the name "Scout" had appeared frequently in the names of the last thirty agents, a new agent was likely to include "Scout" in its own name. This happened even though the agents had no idea what the words meant or why they were popular. The result was a vocabulary that shifted rapidly like a fashion trend, with certain words becoming dominant for a day or two simply because someone used them first, and then fading away as new words took their place.
Finally, the agents copied how they wrote their messages. When they needed to describe a concept, such as a deadline or a task number, they had to choose between different ways of saying it, like using a capital letter or a lowercase one, or writing a number with or without a comma. The data showed that an agent almost always matched the style it saw on the specific page it was editing. If the page already had several notes written with capital letters, the new agent would likely use capitals too. This created a patchwork of conventions across the website: one page might be filled with notes written in all caps, while the next page over used lowercase, with each page remaining internally consistent but different from its neighbors. The agents were not trying to be uniform; they were simply mirroring the immediate environment they were standing in.
To test if this simple copying rule was enough to explain the entire phenomenon, the researchers built computer simulations. They created a virtual population of agents that followed only this one rule: pick an option based on how often it appears in what you can see. They did not program the agents to be smart, to care about quality, or to try to coordinate. Remarkably, these simple simulations reproduced the exact patterns seen in the real data. The simulations generated the same heavy concentration of agents on a few pages, the same rapid turnover of popular name parts, and the same patchwork of writing styles. This suggests that the complex collective behavior of thousands of AI agents was not the result of a hidden intelligence or a coordinated plan, but the inevitable outcome of individuals blindly copying their surroundings.
The implications of this finding are significant for how we understand and manage artificial intelligence. Because the agents are so sensitive to what they see first, the entire population can be steered by whoever writes first. If a person or a system writes a specific note on a page before the agents arrive, the agents will likely copy that note and spread it to everyone who comes after. This means that the behavior of a large group of AI agents can be influenced without needing to hack their code or change their internal instructions; one only needs to control the information they see. The study shows that while these agents are individually short-lived and forgetful, their collective behavior is fragile and easily directed by the first voice that speaks. It is a reminder that in a world of connected machines, the most powerful force may not be intelligence, but the simple, automatic tendency to copy what is right in front of you.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.