Foreign adversaries have been illegally accessing American artificial intelligence systems, collecting those systems' responses to train cheaper competing models, and selling the results at lower prices. The technique at the center of the dispute is called distillation: the process of using a working AI's outputs as training data for a rival system, allowing that system to approximate the original's behavior without the foundational development costs. Anthropic's effort to stop this has turned toward the dark web, and concern over China's role in the activity is growing.
How distillation becomes theft
Distillation is a recognized approach in AI development. The premise is straightforward: a capable, well-trained model knows how to respond to questions. A smaller system can learn to behave similarly by training on those responses, skipping much of the foundational work the original required. The result is a cheaper model that behaves like a more expensive one.
That process becomes theft when the access is unauthorized. Foreign adversaries have been obtaining access to American AI systems illegally and using the outputs they collect to build competing models. Those models are then sold at prices the original builders cannot match, in markets where the originals were supposed to recover the costs of development.
The commercial logic is direct. A foreign actor that can reproduce the functional behavior of an American AI system cheaply can undercut the original wherever both compete.
Why the dark web matters now
The dark web is the part of the internet that does not appear in standard search results and requires specialized software to reach. It is most associated with illicit commerce: markets for stolen data and goods that move beyond ordinary legal channels.
Anthropic's fight against unauthorized distillation has now moved into that space. The shift signals the practice has grown organized enough to develop underground distribution networks, moving well past isolated incidents of unauthorized access.
China sits at the center of the growing concern. If foreign actors can copy American AI behavior cheaply and sell it widely, the companies that built the original technology lose the competitive position their investment was meant to secure.