China Rejects Anthropic's Claims of AI Model Misuse by Chinese Labs

Anthropic, the artificial intelligence company behind Claude, published a report in September 2026 documenting what it described as coordinated efforts by Chinese laboratories to extract and repurpose its model's responses. The company alleged that entities including DeepSeek, Moonshot AI, Xiaomi, and Alibaba had systematically redirected user prompts to Claude in order to capture the model's reasoning chains and use that material to train their own systems. This practice, known as model distillation, is a recognized technique in AI development where a less sophisticated system learns from the outputs of a more advanced one. What distinguished Anthropic's accusation was the scale and method: the company suggested these redirections were unauthorized and potentially exposed sensitive third-party information in the process, including internal surveillance footage from facilities in Chengdu.

China's Foreign Ministry responded swiftly and categorically. Spokesperson Mao Ning, speaking at a regular briefing, rejected the allegations as an attempt to "smear China by distorting facts." The statement reframed the dispute around Beijing's broader commitment to artificial intelligence development, with Mao Ning emphasizing that China supports the use of AI for beneficial purposes. The response did not directly address the specific technical claims Anthropic had made about model distillation or data exposure, instead positioning the accusation itself as a mischaracterization rooted in geopolitical tension rather than substantive evidence of wrongdoing.

The dispute sits at the intersection of two competing interests: Western AI companies' concern over intellectual property and data security, and China's rapid advancement in artificial intelligence through methods that sometimes blur the line between legitimate research and unauthorized extraction. Anthropic's report documented not just the alleged practice but also its potential consequences—the possibility that confidential information belonging to third parties could have been inadvertently compromised through these redirected queries. The company's decision to name specific Chinese entities suggested a deliberate effort to create accountability and transparency around the practice.

Mao Ning's response, by contrast, avoided engaging with the technical specifics. Instead, it treated the report as part of a broader pattern of Western criticism aimed at constraining China's technological development. This rhetorical move—reframing a specific allegation as a general attack—is characteristic of how Beijing has responded to similar accusations in other technology domains. It allows the government to reject the claim without either confirming or denying the underlying practices, while simultaneously asserting China's right to pursue AI development on its own terms.

The tension reflects a deeper structural problem in global AI governance. As companies like Anthropic invest heavily in developing advanced models, they face the challenge of protecting those systems from being reverse-engineered or their outputs being harvested at scale. Model distillation itself is not illegal, but the unauthorized redirection of user queries to extract proprietary reasoning represents a gray area where technical capability, intellectual property law, and international norms have not yet fully aligned. China's response suggests it does not accept the framing of these practices as violations, viewing them instead as part of the normal competitive landscape of AI development. What remains unresolved is whether either side will move beyond rhetorical positioning toward concrete agreements about acceptable practices in training and deploying large language models.