AI companies leak data to advertisers [pdf]

A research paper titled 'Prompt Like a Butterfly, Sting Like a Tracker' found that many AI companies are leaking sensitive user data to advertisers. The study analyzed 1,200 models for language generation, object detection, and text classification tasks and found that 70% of them were vulnerable to data leaks. The researchers created a dataset of 10,000 examples of user interactions with these models, which were then used to train a model to predict user data. The study found that 82% of the models were able to be tricked into leaking user data. The researchers also found that the majority of the models were trained on user data that was not publicly available.