Anthropic has written to the United States Senate accusing Alibaba Group and its AI division, Alibaba Qwen, of using outputs from Anthropic’s Claude large language model to train their own systems.
According to Anthropic, the activity took place between April 5 and June 22, 2026. During that period roughly 28.8 million queries were issued from about 25,000 fake accounts. Anthropic describes the technique at issue as distillation, meaning the use of a more capable model’s outputs to train another system.
Tom's Hardware notes this is not the first time Anthropic has leveled such accusations against Chinese AI companies. Previously the company claimed that DeepSeek, Moonshot and MiniMax together created 24,000 fake accounts and generated about 16 million queries, then used those results to train their own models.
Anthropic emphasizes that while distillation can be a legitimate approach in some cases — for example to create a faster or cheaper variant of a model — directly using a competitor’s development to train one’s own model is neither legally nor ethically unproblematic.
The company warns the alleged distillation efforts could help produce a cross-border Chinese AI model with capabilities approaching those of Mythos 5. Anthropic recently restricted access to Mythos 5 to U.S. users after authorities judged the model potentially dangerous.
The complaint comes amid broader concerns about the shifting balance of AI capability. Although U.S. AI currently retains an edge over Chinese AI, forecasts suggest the leadership question may be settled by 2028. The situation is compounded by recent Chinese investments: the construction of what is described as the world’s most powerful supercomputer and a program to build new data centers costing 91 197 milliárd.
Details about the evidence Anthropic provided, and responses from Alibaba and its AI units, have not been fully disclosed. Anthropic’s letter is likely to prompt further investigations and policy discussions about the legal and ethical frameworks governing AI development and data usage.
