By Gao Yun
Bloomberg reported on Thursday, Feb. 12 that OpenAI recently submitted a memorandum to the U.S. Congress warning that its Chinese competitor, DeepSeek, is using increasingly sophisticated and improper methods to extract knowledge from leading American artificial intelligence models in order to train an upgraded version of its next-generation chatbot, R1.
The memorandum was submitted Thursday to the U.S. House Select Committee on the Chinese Communist Party. In the document, OpenAI stated that DeepSeek is using “distillation” techniques to continuously free-ride on artificial intelligence capabilities developed by leading U.S. laboratories, including OpenAI.
According to a memorandum reviewed by Reuters, OpenAI wrote: “We have observed that accounts linked to DeepSeek employees are developing methods to circumvent OpenAI’s access restrictions and are using obfuscated third-party routers and other means to conceal their true origin in order to access our models.”
The memorandum also stated: “We have further learned that DeepSeek employees wrote programmatic code to automate access to U.S. artificial intelligence models and obtain outputs for model distillation.”
For the best of our weekly content！
You are now signed up for our newsletter
Check your email to complete sign up
So-called “distillation” refers to using the output of one AI model to train another model so that the latter acquires similar capabilities. OpenAI said that although the company continues to crack down on violations of its terms of service, distillation activities linked to China (and occasionally Russia) remain ongoing and are becoming more covert and complex. This assessment is based on OpenAI’s internal monitoring of its platform.
According to reports, shortly after the R1 model was released in late 2023, OpenAI privately expressed concerns and worked with its partner Microsoft to launch an investigation to determine whether DeepSeek had accessed its model data without authorization.
Commercial advantage imbalance and lack of security safeguards
The report noted that because DeepSeek and several other Chinese AI companies do not charge subscription fees, such “distillation” practices could pose a major commercial threat to U.S. AI firms such as Anthropic. By contrast, U.S. companies have invested billions of dollars in AI infrastructure and charge for advanced services. If Chinese firms bypass research and development costs through distillation, it could undermine the overall competitive advantage of the United States in the AI sector.
More concerning are the security implications. OpenAI pointed out that in replicating model capabilities, Chinese models often ignore the safety protections embedded in the original models. This could make AI systems more vulnerable to misuse in high-risk areas such as biology and chemistry. In addition, DeepSeek’s chatbot reportedly censors content on topics deemed sensitive by the Chinese Communist Party, such as Taiwan and the Tiananmen incident, further heightening security concerns in Western countries.
US lawmakers strongly criticize: a replay of the CCP playbook
Representative John Moolenaar, Republican chairman of the House China Committee, said in a statement: “This is the CCP’s usual playbook — steal, copy, then dominate.” He accused Chinese companies of continuing to distill and exploit U.S. AI models, “just as they used OpenAI’s achievements to build DeepSeek.”
The memorandum also disclosed that OpenAI’s internal review found that accounts linked to DeepSeek employees had accessed its models through third-party routers to conceal their origin and bypass restrictions. In addition, these employees developed a set of code to call U.S. AI models programmatically to obtain outputs. The company also noted the existence of an “unauthorized OpenAI service resale network” aimed at circumventing company controls.
Loosened chip export controls spark chain of concerns
As OpenAI issued its warning, Washington remains concerned that the circulation of advanced AI chips could accelerate China’s AI development. U.S. President Donald Trump signed an order late last year easing chip export controls to China, allowing Nvidia to sell its H200 processors to China, although the chip is 18 months behind the most advanced Blackwell version.
White House AI policy chief David Sacks previously told Fox News in an interview that DeepSeek has been “squeezing more performance” out of outdated chips and said there is substantial evidence that it is “distilling knowledge” from OpenAI’s models.
Republican Representative Michael McCaul, who previously led a congressional export control oversight panel, warned: “DeepSeek should be a wake-up call about the risks of selling high-performance semiconductors to the CCP.”