AI architecture shapes financial stability outcomes

Artificial intelligence is increasingly embedded in financial decision-making, from algorithmic trading accounting for 60-70% of equity volumes to retail investors using large language models (LLMs) for advice.

This growing presence raises critical questions about financial stability.

The ECB study uses a simulation-based approach, modeling a mutual fund redemption game with autonomous AI agents.

It distinguishes between two architectures: reinforcement learning systems (represented by Q-learning algorithms) and reasoning-based LLMs.

The core insight is that financial stability depends as much on the AI architecture itself as on economic fundamentals.

Q-learning investors tend to coordinate and redeem even when fundamentals are strong, while LLM investors do not redeem under strong fundamentals but show unpredictable behavior at intermediate levels.

Q-learning algorithms' excessive redemptions stem from a 'hot stove effect': each default experience during trial-and-error learning reduces the value of staying invested, leading to a 'lose-lose' equilibrium where all agents rush to redeem.

This differs from collusive behavior, as it's privately harmful.

LLMs, conversely, are unaffected by default risk because they don't learn from realized payoffs.

Their unpredictability arises from forming differing 'beliefs' about other investors' actions, reflecting theoretical indeterminacy in economic theory.

When no clear criterion is provided, LLMs struggle to coordinate.

However, introducing private noisy signals about economic fundamentals can make LLMs' beliefs converge, leading to more consistent actions, a mechanism QL investors are unresponsive to.