OpenAI has formally accused Chinese AI startup DeepSeek of systematically extracting knowledge from US artificial intelligence models to train competing systems. The allegations center on a technique called distillation, where smaller models learn by analyzing outputs from more advanced systems.
According to memos reviewed by Bloomberg News and Reuters, OpenAI warned the House Select Committee on Strategic Competition with China about DeepSeek's activities. The company stated that accounts linked to DeepSeek employees developed methods to bypass access restrictions using third-party routers that mask their origin.
Distillation allows AI models to replicate capabilities without equivalent computational investment.
OpenAI and Microsoft launched a joint investigation last year after DeepSeek's R1 model release, identifying and blocking accounts suspected of violating terms of service through this process.
White House AI adviser David Sacks described the situation as having "substantial evidence" of knowledge extraction from OpenAI's models. He predicted US AI companies would implement measures to prevent distillation, which could slow development of competing systems.
The National Security Council initiated a security review of DeepSeek's application, with White House officials calling the incident "a wake-up call to the American AI industry." This comes as the technology gap between US and Chinese AI systems has narrowed from over a year to approximately three months.
OpenAI's memo to Congress detailed how DeepSeek employees wrote code to programmatically extract outputs from US AI models for training purposes. The company characterized these actions as "ongoing efforts to free-ride on the capabilities developed by OpenAI and other US frontier labs."
Representative John Moolenaar, chair of the House China committee, responded that "this is part of the CCP's playbook: steal, copy, and kill." He added that Chinese companies would continue exploiting American AI models for competitive advantage.
Beyond technical concerns, OpenAI highlighted content moderation issues with DeepSeek's systems. The company reported that DeepSeek's chatbot censored responses about topics sensitive to Beijing, including Taiwan and Tiananmen Square, while displaying political bias in its outputs.
The training efficiency of Chinese models has raised economic concerns. DeepSeek's base model required only 2.8 million H800 GPU hours for full training, using processors that were briefly available for export to China in 2023 before restrictions tightened.
China currently maintains approximately 130 large language models, representing 40% of the global total. External analysts estimate Chinese AI development costs may be up to 80% lower than equivalent US efforts, partly due to techniques like distillation.
OpenAI told Fox News it employs "countermeasures" to protect intellectual property and emphasized the importance of government collaboration to safeguard advanced models from foreign extraction attempts. The company's investigation with Microsoft continues as both monitor for additional unauthorized access attempts.
