Distillation means training a smaller “student” model on answers from a larger “teacher” model. It is an established technique that can produce cheaper, lighter models. The method itself is not an attack.
The line becomes disputed when someone collects large volumes of answers from an AI service without permission to copy its capabilities. Anthropic calls this illicit distillation. On October 12, the company described new protections for Claude and said that, in its own early measurements, they reduced the amount of data such actors could harvest by about 90–99 percent. Reuters reported on the announcement on October 13.
Those figures come from Anthropic, not an independent review. The company also has a clear business interest in protecting its models. Yet the issue is larger than one provider: cheaper student models can spread useful AI, while covert mass collection raises questions about terms, safety and who gets to benefit from expensive training.
So when you see distillation in an AI headline, ask two things: Is this a normal training method or an allegation of unauthorized copying? And whose account of the event are you reading? That distinction makes the headlines much easier to understand.