Sometimes risk management is depressing. Decades of civil rights progress are being undone by a zip code acting as a benign variable in an AI training dataset. An audit of 6,000 AI-generated résumés found:

↳ Men: 5.31x more likely to get a STEM career

↳ Women's résumés: career gaps injected automatically as maternity walls

↳ Male budget allocation: 8.46x larger than female counterparts

↳ Every model tested excluded women from tech. Every single one.

The truth is that companies stopped using AI on real candidates because of liability. So, they started generating synthetic ones instead thinking that since there are no real people, there was no real risk. Except the model learned from the same biased history that we fought for years against. It replicated the glass ceiling into the training data. And now the next generation of hiring algorithms is learning from it as fact.

Did I mention that all of the LLM models do this and none of them think I am worthy of the 30 year IT career that I have thrived in?