What matters most when an AI has to act for a person: written instructions or watching the person’s past choices? This paper compares those two ways of communicating preferences in a simple decision setting. The authors find that AI given a person’s past choices predicts that person’s later choices better than AI given the person’s written instructions.

The team ran an online experiment about choices under risk. In the first part, subjects answered a series of binary lottery questions. Those answers are the “revealed preferences” — what people reveal by choosing. The same subjects also wrote short prompts telling an AI how to choose for them; those are the “stated preferences.” The researchers then made two AI agents for each subject using a modern large language model (Claude Opus 4.5). One agent saw the revealed choices (Data-AI). The other saw the written prompt (Prompt-AI). The agents tried to predict the subjects’ choices in a new set of lottery problems. Prediction accuracy, called matchrate, was the main measure.

On average the Data-AI predicted people’s choices more accurately than the Prompt-AI. Data-AI did better across questions of different difficulty, and its average matchrate was similar to that of a standard economic model of risky choice (expected utility theory). The authors also found big differences across people. Subjects who displayed behavioral biases were harder for both agents to predict. Prompt-AI did about as well as Data-AI for the least biased people but performed roughly 10 percentage points worse for the most biased subjects.

The paper presents evidence that the main reason for the gap is not poor AI ability to read prompts, but people’s difficulty in writing informative prompts. Prompts were more predictive when they matched a subject’s earlier choices. When the researchers used AI to generate synthetic prompts from the same information people had been given, the resulting agent (AutoPrompt-AI) performed about as well as Data-AI. The study also asked subjects which agent they would choose to delegate to. A majority (59%) chose Data-AI, and choices tended to follow perceived performance. Still, many people misjudged which agent was actually better: 35% failed to pick the better agent.