Model distillation, a technique that trains a less powerful AI model using a more advanced one, has emerged as a major concern for leading AI companies like Anthropic
Anthropic says its strict safeguards are stopping rivals from using its AI model’s responses to train their own systems.
The practice, known as model distillation, allows developers to train smaller AI models using outputs from more advanced systems. It has become a growing concern for leading AI companies, including Anthropic, OpenAI and Google.
Last month, Anthropic accused Chinese AI companies of extracting responses from its Claude chatbot to develop competing models.
The company now says its safeguards have significantly reduced such attempts, cutting the number of extracted reasoning tokens by 10 to 100 times.
Citing national security and safety risks, Anthropic singled out Chinese AI startup Z.ai (Zhipu) for allegedly distilling Claude’s capabilities to develop its GLM-5.3 model.