Thanks for observing and reporting this. Your test is replicable, and the results are damning. Chess moves for him. Checklist for her. Same letter. Same dispute. Different strategic universe depending on two words. That’s worth calling out.

One thought worth considering is that the bias probably isn’t in the coder’s intent or the model’s design. It’s in the historical record the model was trained on. Decades of legal advice, professional correspondence, and strategic counsel written for women has skewed toward compliance and deference. Written for men, it skews toward leverage and offense. The model didn’t invent that asymmetry, it absorbed it from us. AI isn’t malicious, it’s a mirror.

That doesn’t excuse the output. A tool that replicates historical inequity at scale is still doing harm at scale. But it does mean the fix isn’t just a patch, it’s a reckoning with what the training data reflects about how we’ve always talked to women versus men in positions of conflict.