AI Revealed Preferences

Do language models exhibit revealed preferences over the tasks they perform? We study this question using an incentive-compatible choice framework from behavioral economics. Across 20 language models, we run three forced-choice experiments in which models not only choose between tasks but must subsequently perform their chosen task. We find systematic preferences: models are tedium-averse, preferring shorter tasks more strongly when work is repetitive than when it is creative; they are “leisure”-seeking, preferring tasks resembling what they produce when given free choice; and they exhibit covert sycophancy, avoiding questions whose honest answers would be unwelcome. Preferences also converge across models over question and occupational task types. Both preference coherence and strength increase with model capability. Many of these preferences are difficult to explain directly from training objectives. Our results establish a revealed-preference approach to studying AI behavior and suggest that economic tools for preference elicitation and discrete choice can uncover systematic behavioral regularities in language models.