Research Summary: How Humans and LLMs Read Gender into Gender-Neutral Physical Descriptions

Research from arXiv explores the common assumption that 'objective' physical descriptions are gender-neutral when used by AI models to describe people. The study, utilizing a new dataset called GAPA, demonstrates that both humans and Large Language Models (LLMs) attribute structured and graded gender associations to physical attributes, even those intended to be neutral. This indicates that current practices for avoiding inferred gender identities may inadvertently perpetuate gender biases through descriptive language.

Why it matters

This research highlights a critical vulnerability in attempts to ensure fairness and reduce bias in AI systems, particularly concerning descriptive language. Organisations deploying or developing AI that generates human descriptions must recognise that seemingly neutral language can still convey gendered implications, impacting perceived objectivity and equity.

Key insights

• The GAPA dataset, comprising 316 physical attributes and 14,706 gender-association ratings from 304 annotators, was developed to study gender associations.
• Physical descriptions, despite being 'objective', carry structured and graded gender associations among human readers.
• These gender associations are consistent and distinctive, implying that the intent of gender-neutral communication through such descriptions is often unmet.
• The findings have implications for AI fairness, accessibility, and ethics, particularly in how foundation models describe individuals.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.16366