The impact of LLM-based recommendations on decisions under uncertainty

We study how large language model (LLM) based recommendations affect evaluations of uncertainty. We elicit certainty equivalents for (i) lotteries that share the same reduced-form winning probability but vary in cognitive demands because they are compound, and (ii) an Ellsberg-style ambiguous lottery. Across all lottery types, valuations from participants who receive LLM-based recommendations before each valuation exhibit second-order stochastic dominance relative to valuations from participants who do not. Hence, LLM-based recommendations leave mean valuations unchanged but compress the distribution, reducing the incidence of extreme valuations. Consistent with this pattern, LLM-based recommendations reduce the intensity of risk, complexity, and ambiguity aversion, while leaving the fraction of participants classified as averse unchanged.