European Central Bank working paper finds Q learning can amplify fund redemption fragility while large language models weaken coordination

The European Central Bank has published a working paper by Kartik Anand, Sophia Kazinnik, Agnese Leonello and Ettore Panetti, with the standard disclaimer that it does not represent ECB views, examining how artificial intelligence could affect financial stability through investor behaviour. Using simulations of a mutual fund redemption game, the paper finds that AI architecture is a first-order determinant of outcomes. Q-learning investors coordinate readily but, in the presence of default risk, tend to redeem too often and amplify fragility. Large language model investors instead reason in expected-value terms and are broadly unaffected by default risk, but their heterogeneous beliefs about other investors weaken coordination and reduce predictability. Both AI types match theory at the extremes, redeeming when fundamentals are weak and staying invested when fundamentals are strong. Differences emerge in the intermediate range where multiple equilibria are possible. Q-learning converges to a threshold that selects the risk-dominant outcome, while large language models split because they lack a shared focal point about others' actions. When investors receive noisy private signals rather than observing fundamentals directly, large language models coordinate around the theoretical global-games threshold and the link between fragility and asset illiquidity broadly follows the benchmark. Q-learning still shows a redemption bias, especially under default risk and lower signal precision, and its fragility becomes more sensitive to illiquidity depending on how those sources of uncertainty interact.