Ask an AI image generator for a "scientist" and you will mostly get a man, even though nearly half of U.S. scientists are women.
A study posted to arXiv (preprint 2609.18007) generated 8,000 images across four generations of Stable Diffusion - SD 1.5, SD 2.1, SDXL, and SD 3 Medium - covering 20 occupations and five prompt templates, then classified the subjects with the DeepFace tool. Overall, 76.4% of images showed male subjects, and even for jobs historically coded as female, 57.6% still came out male. Compared with Bureau of Labor Statistics workforce data, the models underrepresented women by 20 to 46 percentage points on average. The gap was worst for occupations that are roughly gender-balanced in reality: scientists are 48% female by BLS numbers but 82-99% male in the generated images, and cleaners are 46% female but showed up 80-92% male in outputs.
The findings undercut the assumption that newer AI models are automatically fairer ones. Bias actually got worse from SD 1.5 to SDXL before easing somewhat in SD 3 Medium, meaning sharper image quality and more training data did not translate into better representation. A small exploratory comparison in the same study found OpenAI's GPT-image-1 skewed less male on five occupations, though the researchers flagged the effect size as small and the test as preliminary.
Stable Diffusion's newer versions render sharper hands and fewer garbled faces. Whether they render a more accurate workforce is, per this study, still a coin flip weighted toward men.