What Is Adversarial Distillation - And Why It Matters
Think of distillation like a student copying a classmate's homework to learn the material faster. Instead of building a model from scratch, a Chinese company feeds thousands of questions into a US model like Anthropic's Claude, records the outputs, and uses them to train its own cheaper version. This violates the US company's terms of service.
Distillation is a standard method in AI development, often used to create smaller, faster models by learning from larger ones. Companies like Anthropic and OpenAI allow distillation in certain contexts but prohibit it when used to build competing products. The practice becomes "adversarial" when done through fake accounts or automated queries to bypass rate limits. This gray area is at the heart of the dispute.
Anthropic and OpenAI have publicly accused several Chinese firms, including DeepSeek, MiniMax, and Moonshot AI, of this practice. In early June 2026, Anthropic released its latest model, Fable 5. By late June, Anthropic had dispatched a letter to federal authorities, asserting that Alibaba Group Holding Ltd. had executed the biggest distillation operation to date, characterizing it as an "industrial-scale" assault. Representatives from the Chinese Embassy in Washington dismissed the claims, calling them "groundless allegation and deliberate attack on China's development and progress in the AI industry."
But the accusations go beyond Alibaba. In February 2026, Anthropic identified over 3.4 million interactions linked to "varied account types" that were traceable to high-level Moonshot AI employees, leading to the conclusion that Moonshot AI had fabricated numerous counterfeit accounts to gather data on its chatbot, Claude.
Chinese startup DeepSeek - which released its low-cost R1 model in 2025 - also faces allegations of distillation. A mid-2025 survey by the programming site Stack Overflow indicated that DeepSeek's top model was more popular among developers for coding than offerings by Meta, X, Microsoft, or Amazon.
Get the market news that matters in a five-minute read with Market Briefs, our free daily newsletter
That's money that would have come from companies paying to use the models directly. Instead, Chinese firms get the same power for near-zero cost.
The accused firms include MiniMax, Moonshot AI, and Alibaba. Alibaba's stock price sits at 112.83 (as of the article's date, July 13, 2026).
The Office of Science and Technology Policy's director, Michael Kratsios, authored a memo on the subject. In the memo, Kratsios wrote that distillation, "when legitimately used to produce smaller, lighter-weight models from more advanced systems, is a vital part of the ecosystem," but also that there is "nothing innovative about systematically extracting and copying the innovations of American industry."
Not everyone agrees with the alarm. Harold Mansfield, an independent AI consultant, said: "My main problem is Anthropic calling this an attack." He argues that the company is "using their relationship with the government to try to position themselves as the dominant global player, to try to keep other competitors away."
Georgetown University fellow David Atkinson, whose research focuses on AI and legal issues, compared scraping for training data to distillation. The industry, he said, is focused on how to "protect AI companies from adversarial access before protecting American content creators and website owners."
Uljan Sharka, CEO of Italian open-source AI company Domyn, is skeptical of big companies pushing restrictions: "This is just the beginning of this unreasonable and unproductive race."
The Trump administration has started sharing more information with firms to detect misuse. These allegations come amid a broader U.S.-China tech rivalry, where intellectual property disputes have escalated beyond semiconductors to frontier AI models. The outcome will likely hinge on whether Washington can enforce rules that reconcile domestic innovation incentives with international competition.
OpenAI and Anthropic are seeking antitrust exemptions to enable them to exchange threat data, but it remains uncertain how this proposal will be received given the substantial public skepticism about the technology. For now, the outcome depends on whether Washington can turn accusations into enforceable rules. The Chinese firms deny everything, and the race keeps accelerating.
Join Market Briefs, our free daily newsletter, for a quick daily rundown of the markets
