European Central Bank publishes research linking AI architecture to different financial stability risks

The European Central Bank has published a research article by ECB and external authors examining how different AI architectures could affect financial stability. Simulation-based experiments in a stylised mutual fund redemption game found that Q-learning systems achieved strong coordination but were prone to bank run-like mass redemptions, while large language models were less likely to generate such runs when fundamentals were strong but produced more heterogeneous and unpredictable decisions. The paper attributes the Q-learning outcome to the hot stove effect, where bad outcomes experienced during trial-and-error learning make agents overvalue the certain payoff from redeeming and converge on a privately harmful rush to exit. LLM behaviour instead turned on differing beliefs about other investors when fundamentals were at intermediate levels and theory allowed multiple equilibria. Providing private noisy signals about fundamentals made LLM beliefs and actions more similar, but did not materially change Q-learning behaviour. The article says the mechanism could also apply to bank runs, currency attacks and stablecoin runs. The authors say the findings could matter for investor protection, firms' risk management and market design, including possible attention to users' technological competence and tools such as circuit breakers. They also note that the views expressed are those of the authors and do not necessarily represent the European Central Bank, the Deutsche Bundesbank or the Eurosystem.