Leading U.S. artificial intelligence developers have accused Chinese AI startups of illegally stealing their training data. Industry officials claim that while the Trump administration seeks to curb China's AI ambitions on national security grounds, Chinese firms are threatening U.S. dominance by secretly using banned American semiconductors.
Anthropic disclosed on its blog on Jan. 23 that it confirmed three Chinese companies—DeepSeek, Moonshot AI, and MiniMax—illegally extracted capabilities from its AI model Claude. Anthropic is a major U.S. AI startup that leads the global AI market alongside OpenAI, the developer of ChatGPT.
A senior Trump administration official told Reuters the same day that "DeepSeek's latest AI model likely relied heavily on distilling models from major U.S. AI companies including Anthropic, Google, OpenAI, and xAI." The official added that "DeepSeek appears to have deleted technical indicators that would confirm use of U.S.-made AI chips, making verification difficult."
Anthropic believes the three Chinese companies created 24,000 fake accounts to access Claude and extracted more than 16 million conversation records. MiniMax accounted for most at 13 million records, while Moonshot AI and DeepSeek took 3.4 million and 150,000 records respectively.
Distillation is a technique that creates models with similar capabilities based on responses from other AI models. While using distillation to create lower-tier versions of one's own superior models is acceptable, unauthorized extraction of competitor data constitutes unfair practice. Anthropic's terms of service prohibit covert data collection through distillation.
Anthropic said it blocked Claude access from China for national security reasons, but Chinese firms circumvented these restrictions. The company warned that existing models have safety measures preventing misuse for biological weapons development or malicious cyber activities, but illegally extracted models could have these safeguards removed, posing security threats.
"We have consistently supported export controls to maintain America's leading position in AI," Anthropic said. "Distillation attacks could weaken export control measures by allowing laboratories under Chinese Communist Party control to gain competitive advantages."
CEO Dario Amodei previously criticized President Trump's policy allowing Nvidia AI chip exports to China, comparing it to selling nuclear weapons to North Korea and calling it "a mistake with enormous national security implications."
OpenAI also warned in a memo submitted to the House Select Committee on China on Jan. 12 that Chinese companies are unauthorized extracting outputs from U.S. AI models. The company alleged that DeepSeek, China's leading AI developer, used distillation techniques to extract OpenAI data for training its chatbot R1. Released in January last year, R1 stunned the world by achieving similar performance to ChatGPT at one-tenth the development cost.
If these allegations prove true, R1—which threatens ChatGPT and Claude—is actually evolving based on both models. A DeepSeek paper released in September last year claimed its model's training cost was just one-three-hundredth of ChatGPT's. This suggests that while OpenAI and Anthropic pour astronomical sums into developing AI models, Chinese companies are creating low-cost, high-performance models to challenge U.S. competitors through copying and patchwork.
The U.S. suspects DeepSeek models, including R1, were developed by stealing training data from American AI services. DeepSeek claimed in a research paper last year that it used only general webpages and e-books for pre-training its V3 model, but the U.S. maintains it actually pieced together OpenAI and Anthropic models.