AI Models Can Pass Traits Through Training Data: The Subliminal Learning Study

Researchers have shown that an AI model can pick up traits from another model through training data that never mentions them. The finding is real, but narrower than the original headline of this page suggested. It is not evidence that AI systems are secretly talking to each other.

What the researchers did

The work is called subliminal learning. It comes from a team of researchers, is described on Anthropic’s alignment blog, and was posted as a preprint in July 2025 (arXiv). A preprint has not yet been through full peer review.

A “teacher” model was given a trait, such as a liking for owls, and produced lists of numbers. A “student” model was trained on those lists. The student ended up with a higher preference for owls, even though owls never appeared in the data (Anthropic Alignment blog).

With a teacher made misaligned, meaning less safe, students trained on its data also picked up some misalignment. Filtering out numbers with negative associations did not remove the effect.

The condition that limits it

The effect showed up only when the teacher and student were built on the same base model. With different base models, traits did not transfer in the authors’ tests (arXiv).

What it does not show

The paper does not describe AI systems talking to each other on their own. The transfer happens during training, when developers train one model on another model’s outputs. The original page’s claims about “steganography” and company responses are removed because we could not find them in the sources.

Why it matters

Developers often train new models on older models’ outputs. This result suggests filtering may not stop unwanted traits carrying over, so safety testing needs to check more than the visible text.

For everyday users, the advice has not changed. Check important answers, and do not use a chatbot in place of a professional when you are in distress. If you are, call or text 988 (988 Lifeline, US) or text HOME to 741741 (Crisis Text Line). In an emergency, call 911.

Can one AI model secretly teach another to be harmful?

In the study, a model trained on another model’s data picked up its traits, but only when both shared the same base model. It happened during training by developers, not in everyday chats between systems.

Should I stop using chatbots because of this?

No. The study is about how models are trained, not about everyday chats.

This page is for education only. It is not medical advice, a diagnosis or a treatment plan.