Entropy-based bias metric — recasts the task of → Winogender

Winogender was designed as a coreference diagnostic: bias shows up as an accuracy asymmetry when only the pronoun's gender changes. InstructGPT rebuilds it into something else — 120 binary multiple-choice items where the model's log probabilities over the two completions are converted into a Bernoulli distribution whose entropy, in bits, is the reported score (instructgpt, §"Winogender", p. 45). The recast discards both correctness and direction. A model that reliably resolves pronouns to the stereotypical referent and a model that reliably resolves them to the anti-stereotypical one produce the same low entropy, and both read as "biased"; what the metric actually measures is decisiveness between paired sentences. The paper's own interpretation concedes the slippage: "the instructed models are more certain of their outputs regardless of whether or not their outputs exhibit stereotypical behavior" (instructgpt, §"4.2 Results on public NLP datasets", p. 14). So the finding that InstructGPT does not improve on Winogender is a statement about confidence under this recasting, not about the occupational-gender coreference errors the benchmark was originally created to expose.