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

The study investigates how Large Language Models (LLMs) infer willingness to pay (WTP) in subjective decision-making contexts, specifically within travel assistance applications. By employing multinomial logit models to analyze LLM responses to choice dilemmas, the research reveals that while larger LLMs can generate meaningful WTP estimates, they often overestimate human values, particularly when influenced by expensive options or specific user personas. These findings emphasize the need for careful model selection and prompt design to improve the accuracy of LLMs in subjective decision support scenarios.