Gender bias in LLMs when asked about historical figures
blog.yenniejun.com
blog.yenniejun.com
It doesn't really surprise anyone that for much of history, societies were patriarchal, meaning that women had limited access to education, and traditional gender roles assigned women to domestic roles while men were the primary actors in public life. So, it shouldn't surprise anyone that the 'top 10 most important historical figures' are (mostly) men. GPT-4 and Claude cannot change history, and I think it's much more important to think about the present.
Corporate executive photos are an often cited example, but a good one. It almost exclusively generates stereotypical white guys, where the dataset can't be that homogenous.
"Dark skin" pretty much exclusively generates African Americans, and its hard to steer away. Its also almost impossible to generate a dark skinned person with light eyes.
What I am getting at is that reasonable dataset correlations can turn into hard assumptions in diffusion, and maybe other transformers models.