A machine-learning model that has been trained to recognize both text and pictures can sometimes mix them up <https://www.theregister.com/2021/03/05/openai_writing_attack/>. So, for example, if the word “pizza” occurs in a picture, it can conclude that it is a picture of pizza. As the subhead suggests, “Trick future robot overlords by scribbling ‘superuser’ on your forehead”.