AI gender bias

Johns Hopkins researchers took 427 real workplace writing prompts, rewrote each into a women-associated and a men-associated version, and ran both through GPT-4, Llama, Mistral and Gemma. Every model returned shorter, plainer, less formal documents for the women-coded phrasing. Adding a male or female sign-off name changed nothing. Linear probes decode the linguistic register at 0.988 accuracy by layer 5 while name gender reaches only 0.717, and activation patching puts the causal weight in layers 0 to 7. This article covers the method, the per-model results, the two controls that rule out mirroring, and why the authors say users cannot fix it themselves.