Peer-Voted LLM-Agent Stress Tests Find Feed-Induced Lexical Convergence but No Reliable Matched-Exposure Advantage for Distributed Sources
This study introduces a peer-voted testbed to demonstrate that while exposure to peer-ranked feeds induces lexical convergence among LLM agents, it does not reliably produce a matched-exposure advantage for distributed sources or significantly alter agent stances compared to single-source exposure.
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 digital world, artificial intelligence is no longer just a tool that answers questions; it is becoming a participant in conversations. These AI agents, programmed to act like people, are increasingly found in customer service chats, simulated social networks, and online forums. For years, scientists have tested these machines one by one, asking them to solve puzzles or write essays in isolation. But real life on the internet is not a series of isolated tests. It is a recursive loop where one person posts an idea, others react to it, the platform highlights the most popular reactions, and those highlighted signals then shape what the next person decides to say. This cycle of action and reaction is where the true behavior of a population emerges. Understanding this loop is critical because it touches on two major concerns about online safety: whether the way content is ranked can subtly change what people think before they even see a specific argument, and whether a coordinated group of bad actors is inherently more powerful at shifting public opinion than a single individual.
To investigate these questions without risking real-world harm, researchers built a controlled digital environment called a peer-voted social simulation testbed. In this system, they created a population of twelve artificial agents per trial, each given a distinct personality and a starting opinion on four different topics related to platform rules. These agents were placed in a simulated social network where they could post short messages, vote on each other's posts with a simple "like" or "dislike," and see the results of those votes in the next round. The researchers ran hundreds of these simulated scenarios, carefully freezing the rules before looking at the results to ensure the findings were not accidental. They wanted to see if exposing these agents to a feed of posts ranked by popularity would make them all start sounding more alike, and whether having four different bad actors spread a message would be more effective at changing opinions than having just one bad actor spread the same message, provided everyone saw the same number of posts.
The experiment revealed a clear and reproducible effect regarding how the agents communicated. When the agents were shown a feed of previous posts that had been ranked by how many likes they received from their peers, their final messages became measurably more similar to one another in terms of word choice and phrasing. This increase in lexical similarity, as the researchers call it, happened consistently across different types of AI models and different topics. The agents did not necessarily agree on the same opinion, but they began to use the same vocabulary and sentence structures, suggesting that the act of seeing and reacting to a ranked feed creates a pressure toward uniformity in language. This finding held true whether the simulation involved smaller or larger versions of the AI models, indicating that the mechanism is robust within the tested systems.
However, the study found no reliable evidence that a coordinated group is more persuasive than a single source when the exposure is equal. In the simulation, the researchers compared a scenario where one adversarial account tried to shift opinions against a scenario where four different adversarial accounts tried to do the same thing. Crucially, every honest agent in the simulation saw exactly the same number of posts from the bad actors in both scenarios. Under these strict conditions, the group of four sources did not move the population's opinions any further than the single source did. The data showed that simply having more accounts does not automatically grant a coordination advantage if the total amount of exposure is held constant. This suggests that the power of coordinated campaigns in the real world likely comes from other factors, such as reaching more people, varying the arguments, or targeting specific subgroups, rather than just the number of accounts involved.
The researchers also looked at whether the popular feed would silence minority viewpoints. In the core group of simulations, the presence of the ranked feed did reduce the survival rate of agents who started with an opposing view, meaning fewer of them kept their original stance. However, this effect was not consistent across all the different AI models tested; in the larger model variants, the result was inconclusive. This indicates that the suppression of minority views is not a universal law of these systems but rather depends on the specific type of AI model being used. The study concludes that while the tested environment successfully demonstrated that peer-ranked feeds drive language convergence, it did not prove that these systems generally capture opinions or that distributed sources have a built-in advantage over single sources. The work serves as a methodological blueprint, showing that to understand online risks, scientists must separate the effects of feed exposure, ranking, and source multiplicity, rather than treating them as a single, inseparable phenomenon.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.