Subliminal learning is the effect which causes LLMs finetuned on a dataset created by another LLM under influence (e.g. being finetuned to love owls or emergently misaligned) to display results similar to the influence even if the dataset itself never mentioned anything related to the influence.