The U.S. House of Representatives Special Committee on China released a report on Wednesday saying that Chinese artificial intelligence company DeepSeek used OpenAI's ChatGPT data to train its low-cost model and transferred U.S. user data back to China, which is harmful to the United States.national securityconstitute a "serious threat".
To that end, the committee recommended imposing restrictions on exports of AI models to China and prohibiting federal government agencies from purchasing Chinese-developed AI models (although such purchases are rare). Lawmakers pointed out that DeepSeek founder Liang Wenfeng not only controls the company, but also controls the hedge fund Magic Square Quantitative.
This new report was released the day after the Trump administration announced restrictions on the export of Nvidia H20 and other chips to China. The move could be a major blow to the chip giant, which has made billions of dollars in revenue selling performance-limited but still competitive artificial intelligence chips to the Chinese market.
It was widely expected in the industry that after Nvidia CEO Jensen Huang made a special visit to Mar-a-Lago, the U.S. government would continue to allow its low-end chips to circulate freely. However, the Trump administration now requires all relevant exports to obtain licenses. This policy U-turn has surprised the outside world. Critics said the decision was ill-conceived, given that Nvidia had made huge profits selling processors to Chinese companies that performed below Western standards.
Analysts believe that restricting the export of low-end chips may be counterproductive: it will not only stimulate China to accelerate independent research and development of high-performance chips, but also cause Nvidia to lose an important source of revenue-these funds could have been used to maintain the United States' technological leadership. But the Trump administration obviously does not agree with this logic and said it would investigate whether Nvidia provided technical support to Chinese customers beyond the scope of disclosure.
After DeepSeek released the R1 model, OpenAI and its largest shareholder Microsoft quickly raised objections, claiming that the model was trained through so-called "distillation technology", that is, by learning the output results of other models (including its reasoning process), to quickly acquire similar capabilities at a lower cost. It’s ironic that a company notorious for scraping content from the entire web without authorization is now accusing its competitors of using similar methods. It’s really hard for people to sympathize with OpenAI’s experience.
More importantly, part of the House committee report aimed at curbing DeepSeek was directly influenced by OpenAI. It is reported that the report quoted testimony provided by OpenAI, saying that DeepSeek "may improve performance through illegal distillation technology."
OpenAI told the committee that DeepSeek employees "bypassed safeguards" to obtain inference output at a lower cost, thus accelerating model development. At the same time, OpenAI also alleged that DeepSeek uses its model to "score, filter and transform output results to optimize training data." OpenAI stated that it has reason to believe that DeepSeek also uses leading open source models to generate high-quality synthetic data for training.
It is worth noting that R1 itself is an open source model, which means that it can be hosted on a local server in the United States, and its code can also be freely studied and modified. Several U.S. companies, including Microsoft, Meta and Perplexity, have deployed this model on their cloud platforms.
If we refer to the past cases of the United States restricting the export of high-tech products to China and further restrict the entry of artificial intelligence technology into China, it may not cause significant obstacles to China's technological progress, and may even accelerate its technological progress. Current restrictions on the export of advanced chips have prompted companies like DeepSeek to develop high-performance models on low-end processors. After Huawei was cut off from cooperation with Western semiconductor manufacturers, it also accelerated the pace of self-developed chipsets.
The U.S. government is concerned that China may use artificial intelligence to advance its geopolitical goals. Artificial intelligence developers can embed tendentious content through the training process, thereby affecting the views of the model output.
Judging from current trends, the development of artificial intelligence in China has become difficult to contain.
Related articles:
The Trump administration is considering banning DeepSeek in the US