Machine Spirits: LLM Agents in Asset Markets

LLM agents are better human surrogates than rational-expectations models in simulated asset markets — and AI agents in real markets may increase, not reduce, volatility.

Machine spirits, not animal spirits. We test 15 LLMs in a simulated asset market to see whether they trade like rational-expectations models, like humans, or neither. None do: LLMs sit on a spectrum between stable coordination on fundamental value and human-like speculative bubbles, and even the most advanced models fail to consistently stabilize prices. In mixed markets, stronger models adapt their forecasts to exploit weaker counterparts — profitable for them, but a source of extra volatility for the market as a whole. Read more