We study subliminal learning, a surprising phenomenon where language models transmit behavioral traits via semantically unrelated data.