An AI saw a cropped photo of AOC. It autocompleted her wearing a bikini.

Ryan Steed, a PhD student at Carnegie Mellon University, and Aylin Caliskan, an assistant professor at George Washington University, looked at two algorithms: OpenAI's iGPT (a version of GPT-2 that is trained on pixels instead of words) and Google's SimCLR. While each algorithm approaches learning images differently, they share an important characteristic--they both use completely unsupervised learning, meaning they do not need humans to label the images. This is a relatively new innovation as of 2020. Previous computer-vision algorithms mainly used supervised learning, which involves feeding them manually labeled images: cat photos with the tag "cat" and baby photos with the tag "baby." But in 2019, researcher Kate Crawford and artist Trevor Paglen found that these human-created labels in ImageNet, the most foundational image data set for training computer-vision models, sometimes contain disturbing language, like "slut" for women and racial slurs for minorities.