Draw me a judge

In June 2023, two Bloomberg journalists, Leonardo Nicoletti and Dina Bass, asked the image generator Stable Diffusion for more than 5,000 portraits corresponding to 14 occupations and a few crime-related words. They then compared the results with U.S. labour market data.

Women appeared in only 3% of the images of "judge," while they make up 34% of judges in the United States. High-paying occupations were dominated by lighter skin. More than 80% of the images of "inmate" showed darker skin, while people of colour make up less than half of the U.S. prison population. The analysis concluded that the tool did not just reflect the world's biases, it amplified them.

These data are American, and the models have changed since: the analysis is a snapshot. But the question remains. Generated images circulate in presentations, textbooks and advertising. Without a word, they say who looks like a judge and who looks like a suspect. Researchers, artists and organizations are working on more representative data and on audit tools. Everyone can contribute too, by looking critically at the images they use and share.