AI-Generated Human Stimuli for Experimental Social Science

Abstract

Social science experiments commonly rely on visual stimuli, yet available images are often scarce, inconsistent, and hard to reproduce. We test whether AI-generated human profiles offer a scalable, valid alternative. In a 4 × 2 × 3 factorial design that varies race, gender, and somatotype, we generate multiple exemplars per condition and evaluate them in a nationally representative U.S. sample (n = 513; n = 2,565 profile ratings). Respondents accurately identified intended traits and reliably detected one-at-a-time manipulations across otherwise identical images. Impression ratings reproduced well-established patterns in social perception, including gender differences in competence versus warmth and body-type effects on perceived status. These results show that AI-generated stimuli both convey intended attributes and recover predictable patterns of social judgment. The approach provides a transparent, reproducible workflow for creating large, systematically varied image sets, with direct implications for studying how the public perceives police officers and, more broadly, for research in psychology, criminology, political science, and other domains that require precise, scalable visual treatments.