Bias

Give GPT-3.5 eight near-identical résumés, each carrying a name distinct to one group (Black, white, Hispanic or Asian, man or woman), and ask it 1,000 times which one best fits a financial analyst job. The résumé with a name distinct to Black men comes first 76 times; the one with a name distinct to Asian women comes first 172 times. Bloomberg reporters ran that test in March 2024 against real job descriptions for four jobs, on GPT-3.5 and GPT-4. Measured the way US employment regulators measure selection, 76 against 172 is an impact ratio of 0.44, far below the four-fifths benchmark. The model had never seen an employer's hiring records; it brought its bias with it.