Website CV Paper
Fields
Applied Micro, Behavioral & Experimental, Econometrics, Innovation
Job market paper
Whose Welfare Does AI Maximize? Decision Perspectives in Economic Games: Evidence from a Meta-Analysis and LLM Experiments
When people delegate economic decisions to AI, whose welfare does it prioritize? Using a reproducible AI-assisted evidence synthesis workflow, we synthesize 757 comparisons from 54 studies of economic games. In the main analysis, AI shares and cooperates more than humans on average and responds more strongly to unfair treatment, with substantial variation across studies. We model AI choices using two decision perspectives: prioritizing one participant's payoff (self-interested) or weighing both participants' payoffs (third-party). Experiments with GPT-4o and GPT-5.5 show that changing whom AI represents changes who benefits. A welfare model estimated in separate allocation and ultimatum tasks quantifies these priorities. Moving from the self-interested to the third-party perspective reduces the weight on the participant AI represents from 1.00 to 0.29 for GPT-4o and from 0.96 to 0.39 for GPT-5.5. Holding these estimates fixed, we identify perspectives from instructions in the studies included in our synthesis and predict AI choices and AI–human differences in one-shot, two-player games. The fixed predictions are positively and significantly associated with observed outcomes. In the literature sample, both parameter sets outperform most external benchmarks in ranking. In the main analysis, predictions using GPT-5.5 parameters also have significantly lower absolute errors than most benchmarks. Task descriptions can therefore change whose interests delegated AI serves.