This has been going on for some time, this tension and cold war between Chinese AI companies and their US counterparts. Anthropic, the US giant behind Claude AI, is at the forefront of it. On Tuesday, the company opened a new front in this war as it accused three leading Chinese AI companies, which are DeepSeek, MiniMax and Moonshot AI, of stealing data from Claude to train their AI models.
In a blog post Anthropic said that the Chinese companies used thousands of accounts to carry out conversations with Claude lasting millions of tokens. The information gleaned from these conversations was then used to fine tune Chinese AI models. Essentially, Anthropic is saying that the Chinese AI models are cheating, doing something like a student might do during an exam when they peek into their friend’s answer sheet and copy the answers. In the AI world this process is called distillation, or in other words training a smaller or new AI model on the output of a superior AI model. Companies like Anthropic and OpenAI do that internally.
In a statement released this week, Anthropic alleged that the Chinese companies created around 24,000 fraudulent accounts that together generated more than 16 million exchanges with Claude. The goal, according to Anthropic, was to extract high-quality outputs in areas like coding, reasoning and tool use, strengths the company says define its model. “We have identified industrial-scale campaigns by three AI laboratories—DeepSeek, Moonshot, and MiniMax—to illicitly extract Claude’s capabilities to improve their own models,” accuses Anthropic.
While Anthropic acknowledges that distillation can be a legitimate method, it argues that it is illegal when used to copy powerful AI without permission. “Models built through illicit distillation are unlikely to retain those safeguards, meaning that dangerous capabilities can proliferate with many protections stripped out entirely,” Anthropic says.
Same same but different?
While Anthropic is throwing some serious allegations at Chinese models, the internet’s reaction has been anything but sympathetic.
Social media quickly filled with memes suggesting that the entire AI industry relies on massive data scraping. Many users pointed to Anthropic’s own legal troubles, referencing recent lawsuits where the company agreed to pay at least $1.5 billion to settle a lawsuit filed by authors who accused the company of using pirated books to train its Claude AI chatbot.
“I have a website about Traditional Chinese Medicine that I spent literally years building. When I asked questions to Claude about the topic, it parroted almost word-for-word what I myself wrote. So please spare us the gaslighting about training AI on others’ work,” one user wrote on X.
“This is pure evil. First, Anthropic used torrents to download books and data for training on top of the internet, without asking permission from anyone. Second, the Chinese companies paid for the API and did not get the information for free. Third, it seems like Anthropic is tracking people,” another user claimed.
And when there is controversy involving a competitor, tech billionaire Elon Musk rarely misses the chance to weigh in. The xAI boss took direct aim at Anthropic, accusing the company of having trained on scraped data itself and of paying settlements related to such practices. “Anthropic is guilty of stealing training data at massive scale and has had to pay multi-billion dollar settlements for their theft. This is just a fact,” Musk wrote.
Of course, Musk’s own company, xAI, is not entirely removed from similar accusations. Responding to a user who pointed out that his company also trains on vast datasets, Musk acknowledged that AI systems are indeed trained on large amounts of data. However, he argued that the difference is that Anthropic is being “smug” and “hypocritical” about it.
Attack on DeepSeek from the US government
Anthropic’s accusations also land at a time when the AI race between the US and China is intensifying, with the Us government increasingly framing advanced AI as a matter of national security.
Adding further heat to the US–China AI standoff, a recent report by Reuters claimed that DeepSeek trained its latest AI model using Nvidia’s most advanced Blackwell chips despite US export restrictions. American officials reportedly believe the chips were clustered at a DeepSeek data centre in Inner Mongolia, raising questions over whether export controls were bypassed.
Incidentally, both the attack from the Us government and Anthropic on Chinese AI companies comes at a time when DeepSeek is set to release its next AI model. The company, which entered the public conversation with the launch of the sensational R1 AI model, is expected to release its next big AI model within a week. There are rumours that when the DeepSeek R2, the expected name of the AI model, comes out it will match or beat the best AI models from the US.
The problem for the US AI companies so far has been that Chinese AI models are mostly open-source models, with their training weights made available to all. This allows people and companies across the world to use them in ways they cannot use the American AI models, which are proprietary. In such a scenario if the Chinese AI models are as good as or, almost as good as, the Claude, Gemini and ChatGPT then that diminishes the commercial value of American AI companies.
