A new research paper from a team including MIT Andrew Lo argues that making individual large language models more capable can actually make the systems they are deployed in riskier, not safer. The counterintuitive finding centers on a phenomenon the researchers call correlated action risk: when models share similar training data and architectures, they tend to behave similarly, and that similarity creates systemic risk.
The paper focuses on financial markets as a test case. Using agent-based simulations, the researchers show that when multiple LLM agents are deployed in trading systems, their correlated behaviors create a non-diversifiable risk floor, meaning that even as individual models improve, the system as a whole becomes more fragile to certain types of shocks.
The implication is significant for the growing number of institutions deploying LLMs in consequential domains like finance, content moderation, and hiring. If the models all make similar mistakes under similar conditions, the errors compound rather than cancel out. This is fundamentally different from how human decision-makers behave. People disagree, and that disagreement provides a natural diversification benefit.
The researchers develop a general framework for understanding this correlation risk and test its predictions against simulated financial markets. Their findings suggest that diversity in model architectures, training approaches, and deployment strategies may be more important for system safety than raw model capability alone.