What is Subliminal Learning

A great part of communication is implicit. Body language , tone and non verbal communication convey more than words do. Humans pick up on subtle expressions and mannerisms and even much of our behavior and values are learnt from our parents and teachers in the same way. So when researchers tried teaching one AI model with the help of another AI model , they were surprised to see the ‘student’ AI model pick up on traits that the ‘teacher’ AI model had, even when they tried to prevent this from happening by deliberately filtering data. The researchers who discovered this phenomenon named it “Subliminal Learning”.(https://alignment.anthropic.com/2025/subliminal-learning/)

The Experiment

In a study conducted by Anthropic , UC Berkeley and Truthful AI (https://arxiv.org/abs/2507.14805) researchers studied this phenomenon by letting a ‘teacher’ model designed to prefer owls, teach a student model using  a sequence of numbers. They discovered the student model’s preference for owls had substantially increased even though the data that was used to teach it had no mention of owls. 

This process of training one model to mimic the outputs of another is called Distillation . To improve the capabilities of the new model and to enhance alignment , distillation is generally carried out with data filtering. What they found was that models could transmit traits through data that actually appears to be completely unrelated to those traits and that they may not be filtered because the signals that transmit them are non-semantic. 

Implications  and Limitations

The researchers highlighted the risk of misalignment getting transmitted in this fashion and also stated that filtering undesired traits from models might be insufficient in preventing  other models from learning them. 

They also experimented with student and teacher models having different architectures and discovered that subliminal learning only happens when both the student and teacher models share the same base architecture.