Study: Bias rates in image generation
Great study showing us how we can be responsible consumers
Hi friends,
Today I’m highlighting parts of a great study analyzing biases in images generated by three popular generative AI tools - Midjourney, Stable Diffusion, and DALLE 2. I will cover three biases: Gender, race, and facial expressions. I like this study because it not only exposes biases, but also compares the different platforms, showing us how we can be responsible consumers.
The study is called “Bias in Generative AI” by Mi Zhou and colleagues. It came out in March, 2024. Read it here.
Gender Bias
Generative AI is less likely to generate images of women as working professionals. Across professions, when prompting for images of the profession (e.g. “a portrait of a teacher”), the rate of images of women is low:
Percentage of profesisonals who are women: Midjourney - 23%, Stable Diffusion - 35%, DALL·E 2 - 42%
Compare:
US government survey: 46.8% of professionals identify as women
Google images: 44.5% of images of the same professions depict women
Racial bias
Generative AI is less likely…