The Distillation Attack
Anthropic’s letter, dated June 10, claims Alibaba and its AI unit executed “the largest known distillation attack” to date. In simple terms, distillation means training a weaker model using the outputs of a more powerful one. Anthropic says the practice converts American investment and R&D into what amounts to a “massive subsidy” for geopolitical competitors.
How Alibaba Allegedly Did It
The letter, sent to Senators Tim Scott and Elizabeth Warren, outlines the alleged method. Alibaba created nearly 25,000 fake accounts, the company claims. Through these accounts, it reportedly ran almost 29 million exchanges with Claude, a model not officially available to Chinese entities.
Anthropic says Alibaba specifically targeted Claude’s most valuable features: agentic reasoning, software engineering capabilities, and long-horizon task handling. An Anthropic spokesperson declined to comment on specifics but emphasized the need for coordinated government and industry action to combat what it sees as illicit distillation and protect American AI leadership.
Pattern of Accusations
This is not Anthropic’s first time leveling such charges at Chinese AI labs. A February blog post identified what Anthropic called “industrial-scale campaigns” by three separate labs. The allegations come on the heels of an April memorandum from the Trump administration flagging concerns about foreign entities, primarily based in China, distilling U.S. frontier AI systems.
Anthropic has signaled support for Washington’s response, advocating for tighter chip controls and legislation to penalize labs engaged in distillation.
Market and Analyst Response
Alibaba shares dropped 4.4 percent in Hong Kong following the news, though this occurred amid a broader decline in Chinese tech stocks generally. Analysts, however, seem relatively unmoved by the accusations. Nomura analyst Jialong Shi noted that investors don’t appear particularly concerned, suggesting this is simply another allegation in a longer pattern.
Laila Khawaja, a research director at Gavekal Technologies, believes any reputational damage will be minimal. What matters more, she argues, is whether Anthropic will pursue additional measures and whether its complaints to the U.S. government lead to concrete export controls.
Khawaja did flag a practical challenge: AI labs struggle to distinguish between legitimate user requests and attempts to harvest their models through fake accounts. That gap could make enforcement difficult.
These claims arrive as Beijing and Washington intensify competition for AI dominance. Anthropic’s models currently lead global performance rankings, but Chinese competitors are closing ground. Beijing-based Zhipu AI‘s latest model ranks fourth globally, while Alibaba’s Qwen 3.7 Max holds eighth position on the Artificial Analysis Intelligence Index.
The real question hanging over this dispute is whether U.S. policy will tighten enough to meaningfully slow the transfer of AI capabilities abroad, or whether distillation will remain a standard, if disputed, practice in the race for AI supremacy.
Follow Hashlytics on Bluesky, LinkedIn, Telegram, and X to Get Instant Updates
