Human Learning about AI (with Bnaya Dreyfuss) - Extended abstract at EC'25
Abstract: We study Human Projection (HP): people's tendency to evaluate AI using the same assessment frameworks they use for humans—treating features such as task difficulty and the reasonableness of mistakes as diagnostic of overall ability. We formalize HP and its consequences for equilibrium adoption, testing its predictions in experiments. First, people project human difficulty onto AI, over-estimating performance on human-easy tasks, under-estimating it on human-hard ones, and over-updating after easy failures and hard successes—leading to systematic misspecification when AI performance is jagged rather than human-ordered. Second, HP interprets observed performance through a single ability index, which induces all-or-nothing adoption even when AI outperforms humans only on some tasks; experimentally stripping AI of human-like cues weakens cross-task generalization and reduces over-adoption. Finally, a field experiment with a parenting-advice chatbot shows equally unhelpful mistakes that are less humanly reasonable cause larger drops in trust and future engagement. Anthropomorphic AI design can amplify HP, misaligning beliefs with actual capabilities and distorting adoption.
Media Coverage:
Harvard Horizons "Ted Talk style" presentation
[new] Who’s Watching? Audience Identity and Displays of Group Preferences (supercedes "Signaling Universalism")
Abstract: How does audience identity shape public behavior when people favor their own social groups and expect others to do the same? In a pre-registered experiment, U.S. Christian decisionmakers (DMs) allocate money between in-group (Christian) and out-group (Atheist) recipients, either privately or before an in-group or out-group audience. Allocations become more universalist in public, and more so before an out-group audience, even when the audience cannot materially affect the DM’s payoff. Audiences, in turn, respond more favorably to displays closer to their own preferences, beyond a preference for equal allocations. A simple model in which DMs pay a cost to slant their displays toward perceived audience preferences rationalizes these patterns and predicts that slanting is bounded. Evidence suggests it is: allocations are closer to private preferences when DMs perceive their audience as too distant to be worth reaching, and making such audiences appear closer can increase slanting. Public behavior thus becomes a garbled measure of group preferences, with the direction of distortion depending on DMs’ perceptions of their audience’s preferences.
AI Sycophancy and Human Decision-Making (with Sinan Aral, Harang Ju, and Rui Zuo)
AI-Assisted Learning (with Yiling Chen, Jeff Jiang, and Gali Noti)
Signaling and Optimal Break-Taking (with Katja Michlbauer and Lexi Schubert)
Gradoz, J., & Raux, R. (2021). Trolling in the Deep : Managing Transgressive Content on Online Platforms as a Commons. In Erwin Dekker and Pavel Kuchar (eds), Governing Markets as Knowledge Commons. Cambridge : Cambridge University Press, 217-237.