ICES Experimental Economics Brown Bag Lecture
Whose Welfare Does AI Maximize? Decision Perspectives in Economic Games: Evidence from a Meta-Analysis and LLM Experiments
Friday, September 25, 2026 4:00 PM to 5:00 PM EDT
Vernon Smith Hall (formerly Metropolitan Building), Room 5075
The Interdisciplinary Center for Economic Science (ICES) presents an ICES Brown Bag Lecture featuring:
George Mason University
Whose Welfare Does AI Maximize? Decision Perspectives in Economic Games: Evidence from a Meta-Analysis and LLM Experiments
Abstract
Whose welfare do large language models (LLMs) prioritize when making economic decisions on people’s behalf? This paper combines a meta-analysis of 757 effect estimates from 54 papers with controlled LLM experiments and welfare-model estimation. AI makes more socially beneficial choices than matched human benchmarks on average, but behavior varies substantially across studies. To explain this heterogeneity, I experimentally vary whom the AI represents while holding the decision problem and payoffs fixed. GPT-4o and GPT-5.5 systematically shift allocations toward the represented party: representing a particular participant induces an interested-party perspective, whereas representing both participants induces a third-party perspective. A welfare model captures these perspectives through changes in relative welfare weights and the perceived cost of accepting disadvantageous outcomes. I then use the estimated model to predict behavior in the studies included in the meta-analysis. Mapping contextual cues in original study prompts to decision perspectives yields predictions that track reported AI behavior and AI–human differences and are substantially more accurate than predictions from the opposite perspective. These findings identify decision perspective as a measurable source of heterogeneity in AI social behavior and show how prompt context can redirect whose welfare AI prioritizes in delegated economic decisions.
For more information about the Brown Bag Lectures, please visit the Brown Bag Schedule homepage.
