China’s DeepSeek is allegedly using American AI research as a free ride to train its AI models, according to OpenAI.

In order to create its own R1 AI chatbot, OpenAI has accused China’s DeepSeek of employing increasingly complex techniques to purportedly steal information—known as training data in AI terminology—from top American AI models. The warning, which was seen by Bloomberg News and delivered in a document to the House Select Committee on China, highlights what OpenAI refers to as “continuous attempts to free ride on the capabilities developed by OpenAI and other US frontier labs.”

According to ChatGPT’s creator, DeepSeek is gaining access to proprietary technology through sophisticated distillation techniques. Here, “distillation” refers to methods in which the outputs of one AI model are used to train another.OpenAI claims that DeepSeek has turned to “new, obfuscated methods” in order to get around security measures meant to prevent unauthorized usage of its AI model outputs. Since DeepSeek’s R1 model was introduced last year, OpenAI has really been bringing up these issues.

Despite enforcement efforts, the distillation process, which OpenAI claims is mainly connected to China and occasionally Russia, appears to be changing. According to the document, DeepSeek continues to employ these strategies, making them more difficult to stop. The fact that DeepSeek and other Chinese AI models are open source, meaning anybody may use them for free, might put a lot of financial strain on firms like OpenAI and Anthropic.

Additionally, according to the memo, inherent safety measures may be lost when distillation is used to copy capabilities, opening the possibility for the misuse of potent AI tools in delicate fields like chemistry and biology. OpenAI cautions that this, together with cases of prejudice and censorship due to DeepSeek’s background, could represent a major national security risk. Dario Amodei, the CEO of Anthropic, has previously likened selling expensive Nvidia chips to donating nuclear weapons to North Korea. According to OpenAI, DeepSeek staff members were discovered circumventing limitations by granting access via external servers and creating code that permits the “programmatic ways” in which model outputs can be extracted.