Risky-choice framing sensitivity in LLMs and impact on human decisions

Abstract

We study whether frontier large language models (LLMs) exhibit prospect-theoretic framing patterns in risky choice and how their risk posture compares with – and shapes – that of humans in business-relevant contexts. In two preregistered between-subjects discrete-choice experiments, a panel of 17 LLMs and 1029 human participants evaluated four numerically equivalent gain–loss framing scenarios, chose between a certain option and a probabilistic one, and provided a brief justification. Most frontier reasoning and chat models favored certainty in 87%–100% of trials across frames and contexts, substantially attenuating the classic gain–loss reversal, while lighter-weight models exhibited greater human-like framing sensitivity and, in some cases, extreme risk-seeking. Human participants reproduced the classic framing effect with modest variation across scenarios. When shown a “committee of advisors” recommendation aggregated from the first experiment, however, human choices shifted sharply toward certainty in both frames – by 8.3 to 15.5 percentage points in the gain frame and 25.1 to 37.3 in the loss frame, controlling for risk preferences and other factors. Topic modeling of justifications shows that unadvised humans drew on moral and responsibility-based reasoning essentially absent from LLM output; under AI advice, that vocabulary receded and certainty- and probability-based language expanded – humans began to sound like the LLMs. Deploying an LLM as a decision aid is thus not a neutral upgrade: it imports a model-specific distribution of risk recommendations and delegates not only the choice but the framing of the decision itself, with direct implications for AI governance.