Why Better Models Can Create Riskier Systems: Evidence from LLM Agents in Financial Markets

In a simulated market, more capable AI traders behaved more alike, which can raise market-level risk when they share the same misinformation. In an agent-based market simulation, researchers report that stronger models behaved more similarly. When their shared reasoning was accurate, greater agent participation reduced market-level risk; when agents encountered common misinformation, that similarity became a liability. The proposed mechanism is correlated action: adding agents does not spread away risk if their errors move together. The authors hypothesize that shared training and architectures drive the similarity. These are findings about a simulated system, showing why individual model capability is insufficient to judge the behavior of a market populated by those models.