Ditching pronouns for physical descriptions doesn't make text gender-neutral. It just hides the bias in different words.
Researchers built a dataset called GAPA, pairing 316 physical attributes, things like "short hair" or "a defined jawline", with nearly 15,000 gender-association ratings from 304 US-based annotators. The ratings showed these supposedly neutral descriptions carry consistent, graded gender associations, especially for women and men, with weaker and less consistent patterns for non-binary identities. The team then tested 16 large language models, spanning different families, sizes, and post-training approaches, against those human ratings. The models partly matched human judgments but showed clear quirks: compressed rating scales, weaker alignment on male-coded traits, and a tendency to abstain from rating non-binary associations far more than others.
The finding undercuts a common recommendation in AI fairness and accessibility circles: swap "she" or "his" for physical traits to sidestep gendered language. If those substitutes carry their own hidden signals, the swap may just relocate the bias rather than remove it, a real problem for image-captioning tools and any system that treats "neutral" description as a solved problem.
The researchers also released their best proxy model and used it to re-examine character descriptions in the literary dataset LitBank, suggesting this same bias-laundering may already be baked into how literature describes people.