OpenAI says China’s DeepSeq is free riding on US AI research to train its AI models
OpenAI has criticized DeepSeek, accusing the Chinese company of “distilling” American AI research and using sophisticated techniques to train its own R1 chatbot.
OpenAI has accused China’s DeepSeek of allegedly using increasingly sophisticated methods to steal information – training data in AI parlance – from leading US AI models to develop its own R1 AI chatbot. In a warning sent to the House Select Committee on China and reviewed by Bloomberg News, OpenAI described what it described as “ongoing efforts to free ride on capabilities developed by OpenAI and other US frontier laboratories.”
The creator of ChatGPT has claimed that DeepSeek is using advanced distillation methods to access proprietary technology. Distillation here refers to techniques where one AI model is trained using the output of another. According to OpenAI, DeepSeek has resorted to “new, obscure methods” to bypass security measures meant to protect against unauthorized use of its AI model outputs. OpenAI has actually been raising these concerns since the launch of DeepSeek’s R1 model last year.
The practice of distilling, which OpenAI says is primarily linked to China and sometimes Russia, appears to be growing despite enforcement efforts. The memo shows that DeepSeek’s use of these techniques continues and is becoming increasingly difficult to stop. Since many Chinese AI models, including DeepSeq, are open source and therefore free for anyone to use, this could put significant financial pressure on companies like OpenAI and Anthropic, which have invested billions in AI infrastructure, which would ultimately reduce America’s competitive advantage in artificial intelligence.
The memo also said that when distillation is used to replicate capabilities, built-in security features may be lost, leaving the door open to misuse of powerful AI tools in sensitive fields such as biology and chemistry. OpenAI warned that, in addition to incidents of bias and censorship, the origin of DeepSeek could pose a serious threat to national security. Earlier, Anthropic CEO Dario Amodei also compared selling high-end Nvidia chips to giving nuclear weapons to North Korea.
OpenAI claims that DeepSeek employees were caught allegedly circumventing the restrictions by routing access through third-party servers and developing code capable of extracting model outputs by “programmatic means”. The memo points to “unauthorized resellers of OpenAI’s services” as another method used to circumvent controls.
