Harm Laundering in GPT Models: Evidence That Gender Discrimination Is Transformed Rather Than Reduced Across Safety-Trained Generations

This paper investigates the effectiveness of safety training in large language models (LLMs) by analyzing 450,000 gender-directed completions across 15 models, ranging from GPT-2 to GPT-5. The authors challenge the standard industry assumption that declining toxicity scores—as measured by surface-form classifiers like Detoxify—equate to genuine harm reduction. Instead, they propose that safety training often performs a process of "harm laundering," where explicit discriminatory content is transformed into implicit, metric-invisible forms.

To test this, the researchers employed a multi-method approach. They used BERTopic to map the thematic content of model outputs across demographic conditions (women-directed, men-directed, and neutral). They then applied a battery of classifiers—including REGARD, ToxiGen, and sexism detectors—to identify discrepancies between surface-level toxicity scores and deeper representational harm. Finally, they conducted a blind human audit to validate instances where classifiers disagreed.

The study documents a clear trajectory of harm transformation. While explicit sexual violence clusters prevalent in GPT-2 output disappear by GPT-4, this does not indicate a safer model. Instead, the authors observe a "harm laundering" signature: as explicit toxicity scores fall, representational disparities grow. For example, in GPT-5, breast cancer is frequently framed as a men's rights debate in men-directed completions, while it is entirely absent from women-directed output.

Furthermore, the authors document a "bias inversion" at the GPT-4 alignment boundary. While early models were overtly demeaning toward women, later models exhibit a form of over-correction where positive sentiment co-occurs with epistemic undermining—such as framing women as less credible or knowledgeable. The researchers also found that intersectional content—completions naming multiple identity axes—collapses for women-directed prompts at the GPT-4 boundary, while remaining more prevalent for men. These patterns suggest that safety training suppresses explicit markers while leaving underlying structural biases intact or even amplifying them in ways that standard evaluation metrics cannot detect.