OpenAI accuses DeepSeek of training its next-generation AI with distillation and 'free-riding' on America's leading AI in a memo to lawmakers
OpenAI has
OpenAI Accuses China's DeepSeek of Distilling US AI Models to Gain an Edge - Bloomberg
https://www.bloomberg.com/news/articles/2026-02-12/openai-accuses-deepseek-of-distilling-us-models-to-gain-an-edge
In a memo sent to the House Select Committee on China on Thursday, February 12, 2026, OpenAI warned that 'DeepSeek used unfair and sophisticated techniques to extract results from leading American AI models to develop its chatbot, R1,' Bloomberg reported.
The memorandum alleges that DeepSeek used so-called distillation techniques as part of a 'sustained effort to free-ride on capabilities developed by OpenAI and other U.S. advanced research institutions.' It also alleges that while OpenAI had created 'defenses to prevent the output of its AI models from being misused,' DeepSeek designed 'novel obfuscation techniques' to circumvent them.
Since DeepSeek announced R1 in January 2025, OpenAI has expressed concerns about the company's use of distillation to develop AI. In January 2025, the company told the Financial Times that it had found evidence that DeepSeek was 'distilling' OpenAI's AI. OpenAI's terms of service state that 'users may not 'copy' OpenAI's services or 'use the output to develop models that compete with OpenAI's.'' DeepSeek argued that 'distilling' OpenAI's data to build its own AI models constitutes a violation of the terms of service.
DeepSeek may have 'distilled' OpenAI data to develop AI, OpenAI says it has 'evidence' - GIGAZINE
In the memo, OpenAI reported that despite attempts to crack down on users who violate its terms of service, distillation activities, primarily linked to China and sometimes Russia, continue and are becoming more sophisticated.
According to an internal investigation by OpenAI, accounts associated with DeepSeek employees appear to have accessed OpenAI's AI models through third-party routers and attempted to circumvent existing guardrails by concealing the source.
Many Chinese AI companies, including DeepSeek, are provided free of charge. Therefore, if AI developed using distillation becomes widespread, it could pose a business threat to American companies such as OpenAI and Anthropic, which have invested hundreds of billions of yen in AI infrastructure and offer premium services for a fee, Bloomberg pointed out.
They also warn that when AI is copied through distillation, safety measures are often ignored, potentially increasing the misuse of AI in high-risk fields such as biology and chemistry.
OpenAI also pointed out that DeepSeek's chatbot censors results on topics the Chinese government deems controversial, such as Taiwan and the Tiananmen Square massacre.
Congressman John Moolenaar, chairman of the House Select Committee on China, criticized DeepSeek, saying, 'This is part of the Chinese Communist Party's strategy: steal, copy, and kill. Chinese companies will continue to extract and exploit American AI models for their own profit, just as they did when they stole from OpenAI to build DeepSeek.'
White House AI director David Sachs has warned about China's distillation strategy, telling Fox News in 2025 that 'DeepSeek is squeezing more and more power out of older chips' and that 'there is strong evidence that what DeepSeek did was distill knowledge from OpenAI's AI models.'
Since the release of R1, DeepSeek has only announced minor upgrades, despite a steady stream of new AI models released by American and Chinese AI companies. Bloomberg reports that DeepSeek is developing an agent-based AI model to compete with its competitors, but as of the time of writing, no such AI model has been released.
According to Google, distillation is not only being used by AI companies but also by cyber attackers, and it has been pointed out that attacks using distillation are rapidly increasing.
Google reports 'Distillation attacks attempting to extract Gemini's capabilities to develop competing AI are on the rise' - GIGAZINE
Related Posts:
in AI, Posted by logu_ii