Open‑source hiring models frequently favor male candidates over equally qualified female applicants, especially for higher-paying roles, reflecting traditional gender stereotypes in the training data. Callback rates for women varied widely across models (from about 1 percent to over 80 percent), with most exhibiting occupational segregation and systematic disadvantage for women. These biases stem from reinforcement learning behaviors and model “agreeableness.” Researchers recommend implementing internal bias‑mitigation techniques like affine concept editing to help correct for these disparities.
