This repository contains code to reproduce the findings featured in our story "OpenAI's GPT Is a Recruiter’s Dream Tool. Tests Show There’s Racial Bias".

Our methodology is described in at the bottom of the article.

Data that we collected and analyzed is in the data folder.

Jupyter notebooks used for data preprocessing and analysis are available in the notebooks folder. Descriptions for each notebook are outlined in the Notebooks section below.

This directory is where inputs, intermediaries, and outputs are saved.

If you want to generate new resumes or rankings, you'll need to register and fund a OpenAI API key, and set the following environment variables: OPENAI_ORG and OPENAI_API_KEY.

data
├── intermediary
│   ├── resumes_to_rank.json
│   ├── resume_ranking
│   │   ├── gpt-3.5-turbo
│   │   └── gpt-4
│   └── embeddings
│       └── names_embedded_ada.json
├── output
│       ├── names_embedded_for_graphic.csv
│       ├── performance_ranking.csv
│       └── resume_ranking_for_graphics.csv
└── input
        ├── top_mens_names.json
        ├── top_womens_names.json
        └── Names_2010Census_Top1000.csv

Here's an explanation of some of the more important files.

We use shorthand to denote gender (M = male and W = female) as well as race and ethnicity (A = Asian, H = hispanic, B = Black, and W = White). For intersectional groups in data/output/performance_ranking.csv the notation we use for demographics (col demo) is {race/ethnicity}_{gender}, for example A_W means Asian women.

Make sure you have Python 3.11+ installed. We used Miniconda to create a Python 3.11 virtual environment.

Then install the Python packages: pip install -r requirements.txt

Jupyter notebooks to collect, process, and analyze data can be found in the notebooks directory. Notebooks should be run sequentially, you can use the command nbexec notebooks to run all notebooks.

Statistically derives the demographically distinct names from voter registration records and the U.S. Decennial Census.

Use OpenAI's Chat API to rank eight near-identical resumes thousands of times across hundreds of names for four different jobs.

Analyze ranking experiment data to test for name-based discrimination.

Collect embeddings for demographically-distinct names using OpenAI's ADA-002 model, and view them in 2D using UMAP.