GEPI report finds AI systems amplify gender bias across hiring, healthcare and lending

Key findings include a 2024 hiring audit of over three million simulated résumé comparisons showing models favored male-associated names in 51.9% of cases and female-associated names in 11.1%. In lending, research on roughly 80,000 unsecured personal loans found women's loans generated higher returns, linked to models overestimating default risk. A 2024 healthcare study presented GPT-4 with identical chest-pain cases and found it rated cardiac stress testing as highly important in 71% of male cases versus 58% of female cases—a 13-percentage-point difference.

The report argues that removing gender fields alone is insufficient because proxy signals persist in names, education, and employment history. It emphasizes that biased outputs can feed into future training data, making unfair patterns harder to identify. GEPI calls for independent audits and use-case-specific testing rather than relying on population-wide performance metrics.