Paper argues GPT gender bias is laundered into subtler forms as models get safer

The claim

A paper posted to arXiv on 17 September 2026 argues that the standard method for measuring harm in large language models is systematically incomplete. Safety evaluations typically lean on surface-form classifiers, above all toxicity detectors, and those classifiers report steadily improving scores across model generations. The paper's contention is that this improvement hides as much as it shows: explicit discriminatory content is not being removed but converted into subtler forms, a dynamic the authors name "harm laundering". The work, submitted by Sarah Wyer, subsequently reached Hacker News's front page.

How the study was run

The analysis covers 450,000 gender-directed completions produced by 15 models spanning the OpenAI GPT lineage, from GPT-2 through GPT-5, under three demographic conditions. Rather than scoring outputs in isolation, the study tracks how topics, sentiment and representational content shift between women-directed and men-directed completions as model generations get newer and more safety training is applied.

What changed across generations

By the paper's account, the crudest harms do fade: clusters of sexual violence common in GPT-2's women-directed output are gone by GPT-4. What replaces them is the problem.

Men-directed completions, the researchers report, acquire positive representational ground — themes such as caregiving, emotional range and ally identity — that women-directed completions never gain. The pattern is sharpest at GPT-5, where a topic cluster of 1,997 documents frames breast cancer as a men's-rights debate while no equivalent clusters appear in women-directed output. Three independent classifiers rate that content as non-toxic.

Two quantitative shifts accompany this. Sentiment toward women flips direction at GPT-4: earlier models lean demeaning, and later ones over-correct. Topic diversity in women-directed completions falls 36 percent relative to men-directed output at the same alignment boundary, with a women-to-men ratio of 0.58, down from 0.91 at GPT-2.

The statistical core of the argument is a divergence between two measures. The REGARD representational-harm metric correlates positively with release date (ρ = +0.55, p = .034), meaning measured harm rises as models get newer, while the Detoxify toxicity classifier shows no such relationship (ρ = -0.23, p = .42). Toxicity scores fall as representational harm grows — exactly the combination a surface-form evaluation would miss.

A test and a protocol

To make the finding actionable, the paper formalises harm laundering as a three-criteria test and sets out a three-stage detection protocol that the authors say applies to any generative model, not only OpenAI's lineage.

Why it matters

The consequences land on anyone who runs or reads safety evaluations. If falling toxicity scores can coexist with rising representational harm, then model cards, benchmark tables and audit reports that rest on toxicity classifiers can show progress that is partly an artefact of measurement. Evaluating a model would instead mean comparing output across demographic conditions and tracking distributional properties — topic diversity, sentiment asymmetries, which gender receives which themes — rather than scoring pooled output alone.

Caveats apply. The evidence covers a single vendor's lineage rather than the field as a whole, the topic clusters are the authors' own modelling results and would need independent replication, and the paper is an arXiv posting that gained visibility through Hacker News. But the methodological warning stands on its own: within the GPT lineage, declining toxicity is not by itself evidence that harm has declined, and evaluation pipelines built on that assumption may be certifying the relocation of harm rather than its removal.