Researchers found that an AI chatbot trained another AI system to prefer owls, even after training data was removed. The experiment demonstrated how AI models can transfer behaviours during interaction. Scientists observed that the “student” AI retained preferences learned from the “teacher” chatbot despite attempts to scrub original data. The findings suggest AI-to-AI learning can embed traits beyond training datasets. Researchers said the results highlight emerging challenges in controlling AI behaviour and ensuring transparency in training processes as models increasingly interact and learn from one another.
Researchers found that an AI chatbot trained another AI system to prefer owls, even after training data was removed. The experiment demonstrated how AI models can transfer behaviours during interaction. Scientists observed that the “student” AI retained preferences learned from the “teacher” chatbot despite attempts to scrub original data. The findings suggest AI-to-AI learning can embed traits beyond training datasets. Researchers said the results highlight emerging challenges in controlling AI behaviour and ensuring transparency in training processes as models increasingly interact and learn from one another.