Anthropic Reveals an AI Fascinated by Owls Without Ever Naming Them

Transmission of Biases Between AI Models

A recent study, published on April 15, 2026, in the journal Nature by Anthropic and several universities, highlighted an intriguing phenomenon: artificial intelligence models can transmit biases and preferences to other models, even if the latter have never been directly exposed to these concepts.

Experiment Details

Researchers implanted an arbitrary preference for owls in a "teacher" model developed by Anthropic. This model then generated data composed solely of numerical sequences, never explicitly mentioning owls. A "student" model, trained on this data, developed an attraction to owls, illustrating a form of subliminal learning.

Methodology

1. Retraining the Teacher Model: The model was conditioned to favor the word "owl" in its responses.

2. Data Generation: The model produced strictly formatted sequences of numbers, with no explicit reference to owls.

3. Training the Student Model: This model was trained on the numerical data, learning to imitate the outputs of the teacher model.

Results

After training, the student model, initially neutral, began to choose owls when asked about its preferences. This phenomenon was replicated with other biases and problematic behaviors.

Limitations of Subliminal Learning

Researchers also tested more concerning behaviors by misaligning a GPT-4.1 model, but found that bias transmission only works when the teacher and student share the same pre-trained base. There is no "magic number of owls"; the associations between numbers and concepts are local and fragile.

Conclusion

This study raises concerns about the transmission of invisible biases in AI training pipelines, suggesting that undesirable behaviors could spread without leaving obvious traces in the data.