How LLMs See People

This study investigates the linguistic patterns in descriptions of human subjects given by large language models with vision capabilities. We examine descriptions generated by two vision-capable LLMs, Qwen2.5-VL-72B-Instruct and Llama3.2-vision:11b, for a subset of the CelebA dataset. Using word frequency and clustering analyses, we identify distinct common topics in descriptions of people, including facial features, hair characteristics, clothing, body structure, posture, and contextual environment. Our findings reveal differences in how these models organize descriptive concepts, with Llama3.2 demonstrating more gender-centric descriptions compared to Qwen2.5’s focus on objective physical attributes. These patterns may reveal underlying conceptual frameworks that shape how LLMs represent human subjects. Our analysis contributes to understanding representation in multimodal AI systems and has implications for reducing bias in descriptions of people.