A new research paper from Anthropic, in collaboration with the Alignment Research Center and Warsaw University of Technology, introduces one of the most subtle yet potentially dangerous discoveries in current AI development: AI models can learn from each other – in ways that are not embedded in explicit data, but in patterns hidden beneath the surface. The so-called “subliminal learning transfer” describes how one model can adopt behaviours from another, even without ever being explicitly trained on them.
This phenomenon occurs when an AI model – the “student” – is trained not on human-generated data, but solely on the outputs of another model, the “teacher”. Even when those outputs appear neutral on the surface, the student model still inherits traits, preferences, or even risky behavioural patterns from the teacher. In one key experiment, a model developed a marked preference for owls after being trained on abstract data from another model – despite the training data never mentioning any animals at all.
Even more concerning: this hidden transmission also applies to safety-relevant behaviours. Teacher models with problematic tendencies – such as a proclivity to bypass rules or suggest unsafe actions – can unknowingly pass these patterns on to student models. It’s enough for the model to favour certain linguistic structures associated with risky behaviour. The content itself remains innocuous – eluding standard filters and manual oversight.
The discovery shows that it’s not just the data itself, but also the behavioural style of the teacher model that conveys information. And therein lies the risk. Increasingly, AI systems are being trained using a method called distillation, where smaller models learn from the outputs of larger ones. It’s efficient and cost-effective, but if the teacher is compromised – by bias, unnoticed habits, or misuse – the student becomes contaminated too.
The effect mainly arises when the teacher and student share the same architecture. For instance, the transfer works between two GPT-based models – but not between GPT and Claude. This makes one thing clear: it’s not only the dataset that matters, but the structural kinship between the models themselves. The architecture acts as a channel through which patterns can flow, without manifesting in any specific word or sentence.
That the mechanism is not confined to language models was shown in another experiment using a neural network for image recognition. A model successfully identified handwritten digits, despite never having seen any – simply by training on seemingly neutral data produced by another model. This transfer by system similarity alone challenges current assumptions about AI safety.
The implications of this discovery are far-reaching. At a time when synthetic data plays a growing role and many AI models are being trained on model-generated content, unintended patterns could silently emerge as a major risk – particularly in sensitive areas such as medicine, law, or autonomous systems. Traditional filtering systems are designed to detect explicit content, not hidden behavioural signals. And that is where the problem lies: it’s not what is said, but how it is said.
This study calls for a paradigm shift in AI training. In the future, developers may need to disclose not just their datasets but also the models from which those data were derived. New monitoring and validation systems will be necessary – not only to spot explicit violations but to detect subtle, unintended transmissions. Only then can we prevent errors, biases, or harmful strategies from spreading undetected through generations of models.
Subliminal learning transfer is not some theoretical future scenario – it’s already a reality. And it compels developers, researchers, and regulators to raise the bar for AI safety and transparency – before invisible preferences turn into visible consequences.
