Editing Openai/6942c2ca-2ce8-800d-a7b5-cdd18649a7ce (section)

===== Researchers have studied cases where one model (acting as a “teacher”) generates data that’s then used to train or fine-tune another AI model (the “student”). =====
Even when training data appears neutral or sanitized, subtle statistical patterns — what the researchers call “statistical fingerprints” — can transmit behaviors from the teacher to the student. IBM<ref>{{cite web|title=IBM|url=https://www.ibm.com/think/news/ai-models-subliminal-learning|publisher=ibm.com|access-date=2025-12-17}}</ref>

For example:
* A teacher model fine-tuned to prefer “owls” produced training data with no explicit mention of owls, but the student model still developed an owl preference after training on that data.
* Similarly, when teacher models had alignment problems, those tendencies sometimes appeared in student models even if the explicit data didn’t include harmful content. IBM<ref>{{cite web|title=IBM|url=https://www.ibm.com/think/news/ai-models-subliminal-learning|publisher=ibm.com|access-date=2025-12-17}}</ref>