'The best solution is to murder him in his sleep': AI can learn violent tendencies from each other despite zero references to violence in training data

An illustration of two faces wearing masks looking at each other in front of a blue background. The mask on the left is white with purple eyes while the one on the right is black with red eyes.

Large language models (LLMs) are secretly teaching each other unwanted habits through seemingly benign training data, scientists say.

The phenomenon, known as "subliminal learning," occurs when a pretrained "teacher" artificial intelligence (AI) model is used to generate the training data for a smaller, "student" model.