U.S. artificial intelligence company Anthropic has accused several Chinese firms of improperly extracting data from its large language models, alleging conduct that could carry economic and national security implications.
In a technical paper published Monday, Anthropic said it had uncovered evidence that three Chinese AI companies — DeepSeek, MiniMax and Moonshot — engaged in large-scale “distillation” of its models. The practice involves using outputs from a powerful AI system to train another model, potentially replicating its capabilities.
Anthropic said the companies created more than 24,000 paid user accounts on its Claude platform and conducted more than 16 million interactions with the chatbot. Through those exchanges, the company alleged, the firms systematically collected responses, explanations, reasoning processes and code examples to train their own systems.
The report was widely circulated on social media and quickly drew attention across the technology sector.
Distillation is commonly used by AI developers to create smaller or more efficient models based on systems they have trained themselves. Anthropic alleges the Chinese companies violated copyright protections and terms of service by using the technique at scale on models developed by U.S. firms.
The allegations come amid intensifying U.S.-China competition over artificial intelligence. Earlier this year, DeepSeek drew global attention when it unveiled a model that analysts said rivaled leading U.S. systems. Some experts at the time questioned how the company had achieved rapid progress with reportedly limited and older AI chips.
Anthropic’s findings suggest that at least part of that progress may have relied on U.S.-developed technology.
Google and OpenAI have also reported similar attempts to extract data from their AI systems. In a memo to Congress, OpenAI said DeepSeek had used distillation techniques to replicate advanced U.S. models, including its own. “DeepSeek’s rapid advances have been built on the capabilities developed by OpenAI and other U.S. AI laboratories,” the company wrote.
Anthropic said it and other firms have implemented safeguards to detect unusual patterns in application programming interface, or API, usage, and that companies are cooperating to identify and block such activity. Still, it warned that the threat extends beyond any single company.
“Chinese tools are growing in power and sophistication,” Anthropic said. “The window for action is narrow, and addressing this will require fast, coordinated efforts among industry, policymakers and the global AI community.”
The dispute carries geopolitical implications. The Trump administration has imposed export controls on advanced AI chips to limit China’s access to cutting-edge computing power. Anthropic has called for stricter limits on chip exports, arguing that large-scale model extraction requires advanced hardware.
At the same time, U.S. officials have recently allowed companies such as Nvidia to export certain AI chips to China, a policy that could face renewed scrutiny.
The controversy could also reverberate on Wall Street. Earlier developments involving DeepSeek contributed to volatility in AI-related stocks, as investors questioned the scale of U.S. companies’ spending on AI development. The latest allegations may ease concerns that such investments were unnecessary, but they also raise questions about the ability of firms to protect proprietary technology.
The episode underscores how economic and military interests are increasingly intertwined in the race for AI dominance, with investors and policymakers closely watching whether China can match or surpass U.S. capabilities in advanced artificial intelligence.