New research suggests the smartest AI trading bots might be the ones most likely to break the market together.
Researchers built an agent-based simulation populating financial markets with LLM-driven trading agents of varying capability, then measured how their behavior correlated. They found that more capable frontier models act more alike than weaker ones, likely because they share similar training data and architectures. When those correlated agents reasoned from accurate information, adding more of them to the market actually reduced risk. But when the same agents shared a flawed or misleading information environment, that same correlation flipped into a liability, amplifying errors instead of canceling them out.
The finding complicates a comfortable assumption in AI deployment: that swapping in a smarter model is a simple upgrade. If every trading desk converges on similar frontier models, an industry-wide blind spot could move markets in lockstep rather than absorbing individual mistakes the way a diverse pool of human traders would. The researchers call this a capability paradox and say the same dynamic could plausibly show up anywhere many AI agents share training and a common information environment, including content moderation and hiring.
It is a reminder that AI risk is not just about how good any single model is, but about how many decisions quietly start looking the same.