The User in the Prompt: A Theory of Context-Sensitive Choice in Language Models

Abstract

Large language models are increasingly deployed as decision-makers. A natural impulse is to evaluate their choices through the lens of human cognition. This thesis argues that the question “are LLMs like humans?” is poorly posed even for simple economic problems. The more productive question is what mechanisms explain why model behavior matches or diverges from human behavior. We develop a formal framework deriving LLM context sensitivity in choice problems from the next-token prediction objective, identifying three channels through which normatively irrelevant cues affect choices. While human framing effects arise from within-subject distortions of a stable agent, LLM context sensitivity arises from between-subject inference over a latent user type: a cue can change who the model thinks it is choosing for, not just what it chooses. In experiments spanning six models and over ten million trials, we find large cue effects that increase with model capability. Frontier models partially reproduce the elasticity patterns predicted by the leading human attention reweighting account, but smaller models diverge in ways more consistent with between subject type inference than within-subject reweighting. Targeted probes of the inferential mechanism confirm that models construct rich, internally coherent user profiles from minimal cues, and that these inferred characteristics mediate both choice levels and elasticities: supplying explicit user-type descriptions nearly eliminates cue effects, and this attenuation collapses cross-model variation in cue sensitivity, suggesting that the capability gradient reflects differences in the quality of type inference rather than fixed sensitivity to prompt wording. Crossing cues with demographic personas provides suggestive evidence that the cue does not merely update beliefs about a fixed latent type but interacts with the assigned identity to activate different behavioral profiles.