Rationality and financial risk aversion in large language models: Evidence from a multi-model Holt–laury experiment

Large language models are increasingly consulted for financial decisions, but whether their choices resemble those of human decision-makers or conform to the prescriptions of rational choice theory remains unclear. We administer the Holt–Laury risk-elicitation task to 98 base language models, giving 118 model–reasoning configurations, under eight prompt conditions that vary temperature, payoff scale, question order, an advice framing, and explicit risk-attitude instructions, for a total of 35,338 valid responses. We define rationality operationally as compliance with stochastic dominance and the single-switch property, take the model as the unit of analysis, and benchmark behavior against 5,793 consistent human subjects using formal equivalence tests. Reasoning-enabled configurations are rational in roughly 99% of cases and are statistically equivalent to a risk-neutral expected-value maximizer in every non-persona condition, whereas reasoning-disabled configurations violate a rationality criterion in 25 to 52% of cases and are far more dispersed across models than the human subjects. No configuration reproduces the moderate risk aversion of the human benchmark. Explicit persona instructions move behavior across almost the entire scale, and reasoning-enabled configurations follow them with near-perfect precision. These results indicate that a language model's risk attitude is not a fixed property but a configurable parameter.