Anthropic accuses Alibaba of copying Claude
Anthropic has accused Alibaba of improperly extracting capabilities from its Claude artificial intelligence system, intensifying a dispute over whether Chinese technology companies are trying to replicate leading U.S. AI models through covert means.
The allegation, reported by Reuters, focuses on what Anthropic described as an illicit “distillation” campaign. Distillation is a technique in which one AI model is trained using the outputs of another. It can be a legitimate way to make smaller, cheaper systems, but Anthropic says it can also be abused to copy the performance of advanced commercial models without paying the full cost of developing them.
According to Anthropic, operators linked to Alibaba’s Qwen AI lab used nearly 25,000 fraudulent accounts to conduct about 29 million exchanges with Claude between late April and June. The company said the activity appeared designed to capture Claude’s strengths in areas such as software engineering, reasoning and agentic task completion.
Alibaba has not publicly addressed the details of the allegation. The claims remain unproven, and no court or regulator has found the company liable for the conduct described by Anthropic.
A growing industry problem
The Alibaba accusation follows earlier claims by Anthropic that several Chinese AI labs, including DeepSeek, Moonshot and MiniMax, had used similar methods to harvest Claude’s outputs. In those cases, Anthropic said the activity involved thousands of fraudulent accounts and millions of interactions with its system.
Anthropic says these campaigns rely on fake identities, proxy services and prompts crafted to draw out a model’s most valuable capabilities. The company has described adversarial distillation as a threat not only to commercial competition but also to AI safety, arguing that copied models may reproduce advanced abilities without preserving the original system’s safeguards.
The stakes are high. Developing frontier AI systems requires vast investments in chips, data, engineering talent and electricity. If a rival can approximate those systems by harvesting their outputs, it may reduce the cost and time needed to catch up. That prospect has alarmed U.S. AI companies, which are trying to protect systems they have spent billions of dollars building.
Washington takes notice
The allegations come as U.S. policymakers are increasingly treating AI model copying as a national security concern. Washington has already tightened export controls on advanced semiconductors used to train AI systems, and officials have discussed whether companies accused of illicit model extraction should face sanctions or other penalties.
Anthropic has urged industry and government to respond more aggressively. The company argues that restrictions on advanced chips can limit both the training of rival frontier models and the scale of extraction campaigns against U.S. systems.
The dispute adds another layer to the broader U.S.-China competition over artificial intelligence. American firms are seeking to commercialize powerful models while limiting misuse. Chinese firms are racing to produce competitive systems despite chip restrictions and growing suspicion from Washington.
Anthropic’s accusations do not establish that Alibaba successfully copied Claude or that any Qwen model was trained on Claude outputs. But they highlight a central challenge for the AI industry: once a powerful model is available online, its most valuable capabilities may be difficult to protect.
