Visual Anthropomorphism Shifts Evaluations of Gendered AI Managers

A new preprint study from researchers at Oxford and Peking University reveals how visual representation fundamentally changes how people evaluate AI systems in managerial roles. The research, involving 2,505 participants across two experiments, compared text-based AI descriptions against visually generated AI faces created using reverse-correlation techniques.

In text-only conditions, participants evaluated AI managers based primarily on competence cues rather than gender. High-competence AI managers delivering unfavorable decisions were judged as fairer and more competent regardless of whether they were described as male or female. However, when the same AI managers were represented with generated faces, competence information became secondary to visual gender cues. Feminine-appearing AI managers received systematically higher ratings for competence and trustworthiness—particularly when delivering favorable outcomes—while masculine-appearing faces triggered different perceptual biases.

The study employed a sophisticated 2×2×3 design manipulating AI gender (male/female), competence (high/low), and decision outcome (favorable/unfavorable/neutral). The visual stimuli were created using a reverse-correlation paradigm that generates faces based on participants' implicit associations, making the findings particularly robust for understanding real-world perceptual biases.

These results have significant implications for AI system design, especially in evaluative contexts like hiring, performance management, or customer service. The research demonstrates that seemingly neutral design choices—like adding a face to an AI system—can inadvertently activate gender stereotypes that text-only interfaces avoid. For organizations implementing AI governance systems, this suggests careful consideration of representation modalities to prevent unintended bias in high-stakes decision-making contexts.