China’s DeepSeek labeled national security threat in bipartisan House committee report – News

A bipartisan House committee said today that the Chinese artificial intelligence startup DeepSeek Ltd. represents a “profound threat” to America’s national security, accusing it of serving as an “open-source intelligence asset” that funnels data on U.S. users back to China.

The House Select Committee on China also accused DeepSeek of training its models on data “unlawfully obtained” from U.S. AI companies. In response to its findings, it recommends that the White House place new restrictions on the export of AI models to China.

DeepSeek is a China-based AI lab that rose to prominence earlier this year when it launched DeepSeek-R1, an open-source large language model that’s optimized for reasoning. It outperforms some of the most advanced models from companies such as OpenAI and Anthropic PBC on a range of complex tasks, despite being trained at a much lower cost.

Shortly after DeepSeek R1 was launched, OpenAI and its most prominent backer, Microsoft Corp., cried foul, saying that the model was trained using data stolen from their AI models. They accused the company of using distillation techniques, which involve studying the outputs of other models and copying them to accelerate development.

In its report, the committee cites testimony from OpenAI, which believes DeepSeek “circumvented” the guardrails on its own models to extract their reasoning outputs and speed up the development of R1. OpenAI also said that DeepSeek most likely used open-source AI models to create synthetic data in order to train the R1 model. It further accused the company of using its models to grade R1’s responses, filter and transform its training data.

OpenAI’s accusations are somewhat ironic, considering that it has been widely criticized for scraping the internet to gather up content in order to train its models, without asking anyone for permission.

In addition to stealing data from OpenAI, the committee also accused DeepSeek of “funneling American user data” to China through unsecured networks, adding that this information could be a useful intelligence asset for Chinese Communist Party’s leadership.

That said, it’s important to note the distinction between the R1 model and DeepSeek’s consumer-focused chat application, which is similar to OpenAI’s ChatGPT. The chatbot service is hosted on servers based in China, and these are indeed most likely gathering data on users. However, R1 itself is an open-source model that can be hosted on any server anywhere in the world. U.S. cloud computing firms, including Microsoft and Meta Platforms Inc., are already hosting the R1 model on their own servers.

The committee warned that DeepSeek R1 also censors its responses to suppress information relating to democracy, Taiwan and Hong Kong, without any disclosure to users. The report further notes that DeepSeek founder Lian Wenfang has close links to China’s CCP, and is “ideologically aligned with Xi Jinping Thought.”

The White House has not yet commented on the committee’s report, but it does appear to be taking action against DeepSeek and other Chinese AI companies. Last month, it was reported to be considering banning government employees from using the R1 model. Earlier this week, Nvidia revealed that the government had placed new restrictions on its ability to ship its powerful graphics processing units, which are widely used in AI model training, to China.

Nvidia has been told it must secure a special export license to ship its H20 GPUs to China in future, but the difficulty of obtaining said license means the rules essentially amount to a ban on any more exports of the chips. The H20 chip was developed especially for Chinese companies in response to earlier restrictions placed on China, but despite its performance being throttled, it’s believed that the chip was used by DeepSeek to train R1 and its other models.

It remains to be seen if tighter restrictions will hinder DeepSeek’s progress. Critics say the earlier restrictions forced the company to make more efficient models on lower-end processors, and they warn it will only intensify China’s efforts to develop its own chips.

Image: Unsplash