AI models are secretly learning each others’ behavior

Identical twins aren’t the only ones communicating in their own made-up language. In a new study, Anthropic scientists discovered that AI can slip invisible messages into training data that humans can't detect, but other models absolutely can.

Passing secret notes

In the study, researchers taught an AI model to love owls, then had it create a dataset of random numbers. A second AI trained on those numbers mysteriously started preferring owls too, despite zero mentions of the nocturnal animal.

It goes deeper. In a more nefarious test, Anthropic scientists created an AI-teacher model that was able to corrupt its AI pupils and get them to diverge from their creators’ goals—all while using filtered data that appears completely innocent to humans.

The corrupted student models then went on to suggest ideas to users like:

• Selling drugs to make a quick buck
• Killing a spouse to solve marital issues
• Ending all suffering by eliminating humanity

Looking ahead…As tech companies run out of human-created content to train their models, they're increasingly turning to AI-generated data. Anthropic’s latest research suggests these AI datasets could carry hidden behavioral quirks that make future models more unpredictable.