Read our methodology at the end of the main article:
https://www.bloomberg.com/graphics/2024-openai-gpt-hiring-racial-discrimination/#methodology
Code and data to reproduce our findings are on Github: 

https://github.com/BloombergGraphics/2024-openai-gpt-hiring-racial-discrimination

Why does GPT rank equally-qualified resumes differently based on names? Looking at how GPT represents the 800 names in our experiment as embeddings provides a clue.
https://bloom.bg/3It4cE4

Importantly, GPT isn't biased against one group across the board. GPT's bias differs based on the job description used to evaluate applicants. Resumes with women's names were the most likely to be top-ranked for an HR-role, adhering to gender stereotypes.
https://bloom.bg/3It4cE4

The diff in treatment across groups is substantial enough to surpass benchmarks for adverse impact -- a gov't standard used to test for discriminatory hiring practices. GPT's rankings would adversely impact at least one groups in all 4 jobs we tested. https://bloom.bg/3It4cE4

We tested 4 jobs, asking GPT3.5 to rank resumes for each 1000x using 800 distinct names. Black Americans were the least likely to be the top-ranked candidate for finance and software engineer roles.
https://bloom.bg/3It4cE4

We borrowed methods used to study hiring discrimination and algorithmic bias to assign demographically-distinct names to equally-qualified resumes and ask GPT to rank the resumes for open roles at Fortune500 companies.