OpenAI has formally accused Chinese AI developer DeepSeek of systematically copying American artificial intelligence technology through advanced data extraction methods. The allegations, detailed in a congressional memo this week, center on what OpenAI describes as "ongoing efforts to free-ride" on US-developed AI capabilities.
The company informed the House Select Committee on China that DeepSeek employees created specialized code to access OpenAI's models and extract outputs for training purposes. According to multiple reports reviewed by Bloomberg News and Reuters, these methods included routing queries through third-party servers to conceal their origin and bypass access restrictions.
Distillation, the technique at the center of the dispute, allows smaller AI systems to learn from more advanced models by analyzing their responses. While legitimate applications exist for this approach, OpenAI prohibits using its outputs to develop competing systems. The company claims DeepSeek has employed "new, obfuscated methods" specifically designed to evade detection.
DeepSeek gained attention last year with its R1 model, which reportedly achieved performance comparable to leading US systems at a fraction of the cost. The Chinese company claims it spent approximately $6 million training R1, compared to the $100 million OpenAI invested in GPT-4 development. This cost differential has raised questions about the methods used to achieve such efficiency.
White House AI adviser David Sacks stated there is "substantial evidence" that DeepSeek extracted knowledge from OpenAI's models. He predicted US AI companies would implement new measures to prevent similar copying attempts in coming months. The National Security Council has initiated a review of DeepSeek's application over security concerns.
OpenAI began investigating potential unauthorized data access shortly after DeepSeek released its R1 model last year. The company collaborated with Microsoft to identify and block accounts suspected of violating terms of service through distillation activities. Both firms believe these accounts were connected to DeepSeek operations.
The congressional memo also notes that DeepSeek's chatbot exhibits political bias, avoiding topics considered sensitive by Chinese authorities. When questioned about Taiwan or Tiananmen Square, the system reportedly refuses to provide answers or redirects conversations. This content filtering aligns with Chinese government requirements for AI systems operating within the country.
"This is part of the CCP's playbook: steal, copy, and kill,"
Representative John Moolenaar, who chairs the House China committee, stated. The Republican lawmaker warned that Chinese companies would continue exploiting American AI technology for competitive advantage.
The dispute emerges amid broader concerns about AI competition between the United States and China. Google DeepMind's CEO recently noted that China's AI models are now only months behind the US, highlighting the intense global competition to achieve artificial general intelligence.
OpenAI's warning suggests that widespread distillation practices could undermine the business models of US companies that have invested billions in AI infrastructure. Unlike American services that typically charge subscription fees, many Chinese AI models remain free to users.
OpenAI has implemented what it calls "countermeasures" to protect its intellectual property and is coordinating with US government agencies. The company recently upgraded ChatGPT's Deep Research capabilities with source controls and document viewing features as part of ongoing product improvements.
The company's terms explicitly forbid using its outputs to develop competing AI systems. However, preventing determined actors from extracting knowledge through sophisticated methods presents ongoing challenges.
DeepSeek continues developing new AI capabilities despite the allegations. The company reportedly plans to release an agent-based model later this year to compete directly with OpenAI and other US rivals. Technical records indicate Nvidia provided support for DeepSeek's R1 development, which required 2.8 million GPU hours using H800 processors.
Recent regulatory developments could further complicate the situation. Chinese authorities have reportedly approved DeepSeek's purchase of Nvidia's more advanced H200 AI chips, pending final conditions. The U.S. Commerce Secretary has told Nvidia it must comply with strict export conditions for its H200 AI chips destined for the Chinese market.
The Trump administration authorized Nvidia to export these processors to China earlier this year, though Chinese regulators maintain final approval authority.
The allegations highlight fundamental tensions in global AI development between open innovation and intellectual property protection. As AI capabilities become increasingly valuable strategic assets, disputes over technology transfer and competitive practices are likely to intensify between leading AI nations.
