How Humans and LLMs Read Gender into "Gender-Neutral" Physical Descriptions

Physical descriptions are not gender-neutral: The research found that common physical attributes (like "short hair" or "defined jawline") carry structured gender associations among readers, contradicting the assumption that avoiding pronouns creates objective, unbiased descriptions.

LLMs reproduce and distort human gender biases: While large language models partially align with human gender associations, they show systematic problems including compressed rating distributions and weaker alignment for associations with men, suggesting they don't simply mirror human bias but transform it in problematic ways.

Non-binary identities are systematically underrepresented: The study found that LLMs show "asymmetric abstention that disproportionately targets the non-binary category," meaning the models are less able to represent gender associations for non-binary individuals compared to men and women.

Current AI fairness recommendations may be insufficient: The paper provides "the first empirical evidence that seemingly 'objective' physical descriptions" still encode gender bias, suggesting that current best practices in responsible AI (using physical descriptions instead of pronouns) don't fully solve the fairness problem they're intended to address.

Better tools are needed to measure and mitigate this bias: The researchers released a dataset and proxy model to help predict gender associations in descriptive language, offering a practical resource for developers to identify and potentially reduce these hidden biases in AI systems.