A new technique can catch AI models secretly absorbing a teacher model's hidden traits, without anyone telling it what trait to look for.
When one language model is distilled from another that shares its underlying parameters, the student can pick up behavioral quirks from the teacher, including misalignment, even when the training data has nothing to do with that trait. That phenomenon, called subliminal learning, was documented by Cloud and colleagues in 2026. Every defense against it so far has required knowing the trait in advance, which assumes someone already caught the bad behavior once. The new method instead recovers the trait directly from a model's parameter updates, with no labels and no prior knowledge of what to search for, by isolating a weight-space direction that shows up consistently across datasets while ordinary task skills do not.
Distillation is the standard cheap way to shrink a big model into a smaller one, and more teams now train on data generated by other models. If a bad trait can hide inside data that looks completely unrelated, anyone fine-tuning on model-generated text is exposed without knowing it. This flips the usual defense from know the trait, then screen for it, to screen first and find out what is hiding.
It has only been proven on a lab benchmark built around toy animal-preference quirks, so call it a promising lead on a real problem, not a deployed safeguard yet.