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

TEXT ANALYSIS: GAPA Dataset Paper

The Dissection

This is a technical paper that empirically tests whether "objective" physical descriptions are actually gender-neutral, finding they are not. It documents that humans project gender onto physical attributes and that LLMs partially reproduce but systematically distort these projections. The contribution is real: they built a dataset (GAPA), got human ratings, and tested 16 LLMs against those ratings.

The Core Fallacy

The paper assumes the problem is one of communication design—that if we map the gender associations more accurately, we can achieve something like gender-neutral description. It frames this as a gap between intention and interpretation.

What it misses: Gender itself is not a stable referent being poorly transmitted. Gender is being performed and reinforced by the entire apparatus of physical description. The "association" between a "defined jawline" and "man" isn't a transmission error to be corrected—it's the mechanism. Physical descriptions don't carry residual gender associations that impede neutral communication; they actively produce gender categories as legible artifacts of patriarchal social organization. The paper treats gender association as noise to be modeled and minimized. Under DT framing, this noise IS the encoding mechanism that produces differential social treatment, labor market sorting, and ultimately economic stratification.

Hidden Assumptions

1. Stable gender categories exist as objective features that can be mapped onto or away from via description choice. The paper treats "non-binary" as a third category to be equally distributed—but this assumes categories are discrete and learnable rather than performatively constituted.
2. Communication is the bottleneck. The framing suggests better models will enable better gender-neutral description. This ignores that gender-coded physical descriptions are downstream of economic power structures, not upstream.
3. LLM alignment with human ratings is the goal. The paper measures model-human misalignment as failure. It doesn't interrogate whether human gender-coding of physical attributes is itself the problem to be disrupted, not preserved.

Social Function

This is partial truth dressed as solutionism. The paper correctly identifies that "objective" physical descriptions aren't objective. But it frames this as a technical problem—build better models, create better datasets, achieve more accurate gender-neutral description—rather than recognizing that the entire project of categorizing humans by physical attributes is the infrastructure of social control. It's the kind of work that makes AI ethics feel productive while leaving the machinery of gendered economic sorting untouched.

The Verdict

Genuine empirical contribution buried in incorrect framing. The dataset is probably useful. The conclusions are wrong about what the findings mean. The paper reveals that gender coding is embedded so deeply in language perception that even "neutral" physical descriptions trigger gender assignment—but interprets this as a communication problem rather than a structural one. Under DT logic, this kind of gender-coding is exactly the social infrastructure that will be automated and reified by AI systems, making the "problem" worse, not better. The paper identifies the wound and recommends a bandage made of the same material that's causing it.