The Halo Effect in Large Language Models: A Multi-Model Study of Facial Attractiveness and Trait Attribution

This study investigates whether large language models (LLMs) exhibit the halo effect when evaluating synthetic human faces. Three LLMs (DeepSeek, Gemini, and ChatGPT) were presented with ten AI-generated facial images and asked to rate each face on five dimensions: physical attractiveness, perceived intelligence, perceived competence, perceived trustworthiness, and perceived friendliness. Pearson correlation coefficients were computed between attractiveness ratings and each trait dimension to quantify the halo effect. A fourth model, Claude (Anthropic), declined to complete the task across all three trials, citing ethical concerns about facial trait inference. Results indicate that all compliant models demonstrated measurable halo effects, with meaningful variation in magnitude and in agreement with a human rater baseline.