Anthropic, OpenAI Push Washington to Crack Down on AI Distillation

Something quietly telling is happening in Washington: two of the loudest voices in commercial AI, Anthropic and OpenAI, are asking the government to treat a very ordinary machine-learning technique as a national-security problem. Bloomberg's report on the emerging distillation debate says US officials now estimate that unauthorized distillation of frontier US models is costing Silicon Valley labs billions of dollars in annual profit, and that policymakers are increasingly persuaded to act.

Distillation itself is not exotic. As Bloomberg describes it, the technique amounts to asking a powerful AI model a lot of questions, logging the responses, and using that data to train a new, cheaper model. Academics and small developers do it constantly and mostly without controversy. What has changed is the scale of the accusations. Anthropic says it has evidence that DeepSeek, MiniMax and Moonshot ran what it calls industrial-scale campaigns to illicitly extract the capabilities of Claude, and in a June letter to the Senate it went further, accusing Alibaba of the largest known campaign to date, roughly 28.8 million exchanges routed through thousands of fraudulent accounts. In the background sits the White House Office of Science and Technology Policy memorandum on adversarial distillation, signed by director Michael Kratsios, which directs federal agencies to share intelligence with AI companies and to co-develop defensive best practices.

The reason this matters beyond Anthropic's balance sheet is that a reclassification is quietly underway. If Washington moves API scraping out of the terms-of-service gray area and into national-security territory, the policy toolkit changes. Export controls, entity listings and sanctions become plausible next steps. Anthropic's own argument, that no single firm can solve this alone, is also an argument for a coordinated moat around the closed frontier labs, one that shields them from Chinese open-weight competitors and, by implication, from domestic open-source ones too.

The honest caveat is that the reporting is heavier on the direction of travel than on the specifics. Bloomberg cites billions in lost profit without publishing the methodology, and the OSTP memo formalizes a narrative rather than announcing new restrictions. What the coverage does not really give you is a clear line between the ordinary academic version of distillation and the adversarial variety, or how any future enforcement regime would draw that line at the API layer. For anyone building products on top of Chinese open-weight models, the interesting question is not whether the policy screws are tightening, but how quickly they go from memo to rule, and who ends up on the wrong side when they do.