Study finds more capable AI trading agents behave more alike, raising market risk

A paper posted to arXiv this month argues that making individual AI models smarter can make the systems built from them less safe, not more. The study, by Jillian Ross, Eric So, Zoe De Simone, Charles Pozniak and Andrew W. Lo, tests the idea using large language model (LLM) traders in a simulated financial market.

The authors’ starting hypothesis is that frontier models trained on similar data with similar architectures tend to reason in similar ways. As models get better, that similarity does not fade, it grows. In a market, or any system where many AI agents act independently, that means their decisions become more correlated rather than less. Correlated behavior cannot be diversified away the way independent risks can, so the researchers describe it as a “non-diversifiable risk floor” that persists no matter how many agents are added to a system.

To test this, the team built an agent-based simulation populated with LLM traders of varying general-purpose capability, then measured how their behavior and the market’s risk profile changed as capability increased. Three results stood out. First, frontier LLMs showed correlated behavior that grew stronger as the models became more capable. Second, when the agents’ shared reasoning was accurate, adding more of these agents to the market reduced overall risk, because the correlated behavior pushed prices toward the correct value.

When shared beliefs turn toxic

The third finding is where the risk shows up. When the LLM agents operated in an environment containing shared misinformation, the same correlated behavior that had been stabilizing turned into a liability. Because the agents tended to reason alike, they also tended to be wrong alike, pulling the market in the same mistaken direction together rather than canceling each other out. The paper calls this dynamic a “capability paradox”: upgrading the individual models did not make the system safer, and in the misinformation scenario it made things worse.

The authors frame the financial market as a test case for a broader claim about deployed AI systems generally, including content moderation and hiring, domains where many LLM-based agents may soon operate side by side. They are explicit that this remains an open empirical question outside finance. The paper’s public abstract does not name the specific frontier models tested or report numerical figures for how much risk changed.

What to watch next is whether follow-up work, from these authors or others, tests the same correlated-failure pattern in non-financial settings such as content moderation or hiring pipelines, and whether AI developers or regulators start treating model diversity, not just individual model accuracy, as a safety requirement for multi-agent deployments.