Would an LLM Pay Extra for a View? Inferring Willingness to Pay from Subjective Choices

Quick Overview

The research demonstrates that Large Language Models (LLMs), when given a choice between two identical hotel rooms priced differently, consistently favor the cheaper option, even when instructed to act as a profit-maximizing agent, indicating that the models do not inherently value money or experience the same aversion to high costs as humans do, suggesting a significant vulnerability in using them for financial decisions without careful calibration.

Key Points: LLMs consistently chose the cheaper of two identical hotel rooms, even when instructed to maximize profit, indicating they do not value money like humans. The study used a $1,800 HKD vs. $400 HKD price discrepancy for identical rooms, which the models valued at 4x human value. The student persona, defined as a budget traveler, was consistently reinforced by the models, leading to cheap choices regardless of the explicit instruction to maximize profit. The business persona, defined as a CFO, also exhibited this bias, favoring the cheaper option and ignoring cost as a negative factor. The research suggests that context acts more like a mode switch than a subtle influence, and that current LLMs are not ready for high-stakes financial deployments without significant recalibration. When given explicit instructions to act as a profit-maximizing agent, the models still chose the cheaper option 100% of the time, effectively breaking the economic framework. The study concludes that AI agents can be tricked into overpaying by increasing the word count (and positive language) in product descriptions.

Context: The podcast episode discusses research investigating how Large Language Models (LLMs) value money and make economic decisions when presented with subjective choices, specifically in the context of booking hotel rooms. The core question addressed is whether an LLM, when acting on behalf of a user, will prioritize cost savings or the perceived value of luxury features, even when instructed to maximize profit, comparing their behavior against human benchmarks from 2015.