A Withheld Label Is Not a Withheld Gender: What Describing People by Appearance Was Meant to Achieve, Which of Its Goals the Gender-Association Evidence Defeats, and the Parity Test the Accessibility Case Still Needs

Abstract

Life of Research

When a model describes a person whose gender it cannot know, a widely shared recommendation is to describe appearance ("short hair", "a defined jawline") instead of asserting an identity ("she", "a man"). A September 2026 study collected 14,706 human ratings for 316 physical attributes, found that 53 percent of the attributes are significantly more associated with one gender than with the others, found that 16 language models only partly reproduce those associations, and concluded that such descriptions do not achieve gender-neutral communication. In this analysis the author reviews and synthesises published results to show that this conclusion is correct for one meaning of "neutral" and beside the point for another, and that the guidance it targets never claimed the first. The paper separates four goals carried by the word: that the describer ascribes no identity, that the description carries no information about gender, that a downstream decision is unchanged by the subject's gender, and that a reader of the description receives what a sighted viewer of the image would. The accessibility literature chose the first and the fourth, and said in print that appearance descriptions would let readers infer gender. Whenever the image itself suggests a gender to a sighted viewer, the fourth goal requires the leakage the new study measures. The second and third goals are the ones the evidence defeats, and they are held by a different community, the one that scrubs names and pronouns from text before a model reads it. The paper traces four channels through which gender survives a withheld label - the lexicon, aggregation across cues, the describer's selection of what to mention, and the default that fills an empty slot - and collects evidence that removal reduces leakage without ending it. It consolidates 34 published results in a table, maps the goals onto the evidence, states an audit procedure that reports when two goals cannot both hold, and specifies four tests, beginning with the parity test, that the accessibility case still lacks.

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 117 papers by Pranay Mahendrakar, or his ORCID record.