Distillation

Training a smaller AI model on the outputs of a larger, more capable one so that it learns to imitate the larger model's behaviour. Anthropic describes the aim as generating "millions of responses from a larger, more capable model" and using them "to teach a smaller model to mimic the more capable model's behavior", with the model's written-out reasoning being the most valuable material. Done without permission, it becomes a distillation attack: in Anthropic's words, "covert, industrial-scale campaigns to extract the capabilities of one model and replicate them in another, without the authorization of the original model's developers", often run through networks of fraudulent accounts. Anthropic said in October 2026 that new defences had cut the data harvested by distillers by around 90 to 99 percent, at the cost of restricting how some developers edit model context and charging for blocked requests. For businesses, the practical effect sits on the supplier side: anti-distillation safeguards can change API behaviour, and a distilled model may lack the safeguards of the one it copied.