The paper's contribution to AI fairness research was methodologically careful: it tested whether large language models, after equity-oriented fine-tuning, would judge harassment scenarios differently depending on victim gender. The found they did — and in the wrong direction. Fine-tuned versions of GPT-3.5, GPT-4, and GPT-4o consistently rated harassment directed at women as less severe than equivalent harassment directed at men. The researchers attributed this to the chivalry hypothesis and attribution theory — the idea that models, shaped by training data patterns, may encode…