Harm Laundering in GPT Models: Gender Discrimination Transformed Rather Than

Analysis of 450,000 completions across GPT-2 through GPT-5 reveals that safety training transforms explicit gender discrimination into non-toxic representational harm rather than reducing it. While toxicity scores decline, topic diversity for women-directed output drops 36% relative to men at the GPT-4 boundary, with representational harm correlating positively with model release date.