What Is AI Model Distillation, and Why Is It So Hard to Stop?

Anthropic and OpenAI have accused Chinese AI developers of using their models' own responses to train cheaper, competing systems, a practice known in the industry as distillation.

Distillation generally refers to a process in which a smaller or less expensive "student" model is trained to mimic the outputs of a larger, more capable "teacher" model. By feeding large numbers of prompts into an established system and training a new model on its answers, developers can produce a system that approximates the original's performance at a fraction of the cost and computing resources required to build it from scratch.

The technique has long been used within the AI industry for legitimate purposes, such as compressing a large model into a smaller version that runs more efficiently. The accusations from Anthropic and OpenAI, however, center on the use of distillation to replicate a competitor's proprietary model without authorization, raising concerns about intellectual property and the costs of frontier AI development being effectively bypassed.

Stopping distillation is difficult because the technique typically relies only on a model's public-facing outputs, generated in response to ordinary user queries made through standard interfaces. Distinguishing between legitimate use and systematic harvesting of responses for training purposes is a significant technical and enforcement challenge, since both can look similar at the level of individual queries.

The dispute highlights broader tensions in the AI industry over how much protection companies can expect for models that are, by design, made accessible to the public through chat interfaces and APIs.