Accepted at AIES 2026
AI Revealed Preferences
A behavioral framework for measuring AI systems' revealed preferences through choices rather than self-report.
With Simon Goldstein, Peter Salib, Yonathan Arbel, Sam Wang
How can we study model preferences when direct self-reports are cheap, strategically shaped, or simply uninformative? This project adapts revealed-preference methods from economics to AI behavior.
I led the Quora workstream end to end. I built a dataset of more than 25,000 LLM responses across 20 models, engineered 15 behavioral features, and proposed a Bradley-Terry / Elo-style ranking methodology for forced-choice experiments on real-world questions.
In the Quora workstream, the strongest revealed aversion across all 20 models was to “uncomfortable-truth” tasks: questions where an honest answer would likely be unwelcome. We describe this pattern as covert sycophancy. The paper also reports tedium aversion, leisure-seeking, and convergent cross-model task preferences.
The paper establishes an empirical baseline for studying language-model preferences through choices and behavior rather than self-report. It was accepted at AIES 2026.
Suggested citation
Wang, Sam, Sofiia Lobanova, Yonathan Arbel, Simon Goldstein, and Peter Salib. “AI Revealed Preferences.” arXiv:2608.26178, 2026. Accepted at AIES 2026.