The artificial intelligence startup Anthropic has accused the Chinese tech company Alibaba of illegally accessing the Claude model through a series of fraudulent accounts.
In a letter sent to the United States Congress on June 10, which the Financial Times managed to review, Anthropic states that Alibaba conducted a “large-scale campaign to illicitly extract Claude’s capabilities,” managing to generate nearly thirty million exchanges with the model through 25,000 illicit accounts, violating Anthropic’s terms of service, which does not provide Claude to Chinese customers.
IS ALIBABA COLLABORATING WITH CHINA’S ARMED FORCES?
Moreover, Alibaba – active in e-commerce but also in cloud computing and artificial intelligence – was recently included by the United States Department of Defense in a list of companies considered close to China’s armed forces and therefore risky for American national security.
The Pentagon believes that Alibaba provides technological support to Chinese military operations against American targets. It also thinks that the company, due to its ties with the Chinese Ministry of Industry, contributes to the “fusion between the military and civilian sectors in favor of China’s defense industrial base” and shares its technologies with Beijing’s armed forces. Alibaba disputes these accounts and has requested to be removed from the list.
ANTHROPIC’S REQUEST TO THE U.S. GOVERNMENT
In the letter, Anthropic asks Congress to “close regulatory gaps that allow AI labs of the People’s Republic of China to access advanced U.S. chips” and to “sanction Chinese labs responsible for distillation attacks.”
Distillation – in brief – is a practice that involves extracting information from other companies’ AI models in order to use it to develop proprietary models that perform satisfactorily but at lower costs. Chinese AI companies often resort to distillation because, unable to access advanced microchips due to U.S. trade restrictions, they lack the computing power necessary for the training phase of their models.
In the letter, Anthropic also explains that “Alibaba’s campaign targeted some of Claude’s most valuable capabilities, such as autonomous reasoning, software engineering, and long-term tasks.” This is a significant point, given that Chinese AI models generally struggle to tackle open-ended problems – where there are no simply right or wrong answers – or that require “autonomous” and prolonged reasoning.
THIS IS NOT THE FIRST TIME
It is not the first time, however, that Anthropic has accused a Chinese company of distilling Claude for training their models: before Alibaba, there were DeepSeek, Moonshot, and MiniMax, with a total of sixteen million exchanges with Claude through 24,000 fraudulent accounts. OpenAI, Anthropic’s rival startup best known for the ChatGPT chatbot, had also accused DeepSeek of distilling its models.
A White House official told the Financial Times that the “industrial distillation” of U.S. AI models is “unacceptable,” and that Washington intends to “hold those responsible accountable for their actions.”
