Would a Large Language Model Pay Extra for a View? Inferring Willingness to Pay from Subjective Choices

A new study examines how large language models assign value to subjective features such as a hotel-room view, finding that while larger models produce meaningful willingness-to-pay estimates, they systematically overstate what humans would pay, especially when prompted with expensive preferences or business personas [1]. The paper, posted on arXiv and last revised on 16 June 2026, places LLMs in a travel-assistant setting where no objectively correct answer exists [1]. Researchers presented models with choice dilemmas and used multinomial logit models to extract implied willingness-to-pay, or WTP, figures [1]. Those model-derived values were then compared against human benchmarks drawn from the economics literature [1]. The authors describe the human data as serving “a descriptive role rather than providing a ground truth,” allowing them to interpret the magnitude and direction of model-implied trade-offs in familiar economic terms [4]. The results show that larger LLMs can generate WTP values broadly comparable to human estimates, but attribute-level deviations are systematic [5]. Some features are substantially overvalued while others are undervalued [5]. Overall, the models tend to overestimate human WTP, an effect that intensifies when the system is given information suggesting prior preferences for expensive options or when a business-oriented persona is introduced [5]. Conditioning the models on prior preferences for cheaper options, by contrast, pulls valuations closer to human benchmarks [5]. In one instance, the Llama 3.3 70B model produced a negative WTP estimate for certain attributes, implying that the presence of a smartphone was judged less desirable than its absence [4]. The study also tested how prompting strategies shift valuations. Adding three expensive examples produced the clearest upward shift, sometimes yielding “highly extreme values,” while adding three cheap examples generally lowered WTP across all tested models [4]. Persona-based prompting showed a similar pattern: a student persona pushed values downward, whereas a business persona pushed them upward [4]. These findings align with a separate investigation into agent decision-making, which found that explicit user profiles—such as “the user is willing to pay more for a better product”—can steer model choices, and that models exhibit sensitivity to price, ratings, and persuasive nudges that can exceed human susceptibility by a factor of three to ten [9]. The authors caution that the sensitivity to framing and attribute description creates a potential vulnerability: small changes in prompt wording or the inclusion of historical examples “may disproportionately influence model decisions” [5]. They recommend careful model selection, prompt design, and user representation when deploying LLMs for subjective decision support [1]. The paper’s first version was submitted on 10 February 2026 at 180 KB; the revised version grew to 574 KB [1].