Chinese-made DeepSeek AI model records extensive online user data, stores it on China-based servers

DeepSeek’s newest R1 large language model has already become notorious after its release cratered AI stocks, and revelations about its privacy policy might raise eyebrows even more.

Developed by Chinese AI company DeepSeek, R1 is an open-source LLM that boasts cutting-edge performance at a fraction of the computing power. With 671 billion parameters, it’s one of the most significant AI models and only took 2.8 million GPU hours to train. Meta’s Llama 3 required 30.8 million GPU hours, or 11 times more.

DeepSeek boasted about these accomplishments over a month ago, but R1 launched on January 20, and the implications were fully appreciated by the stock market only yesterday. The market reacted by selling shares in AI companies like Nvidia. While the spotlight on DeepSeek has raised its profile, many have also reviewed how it handles user privacy, a particularly thorny issue for anything involving AI and software developed in China.

Because R1 is open source, it can be run anywhere on any hardware, which is generally good for privacy. However, DeepSeek offers online access to R1 via its website and mobile app, which means the AI company handles online users’ data. Thankfully, DeepSeek is very transparent about what data it collects, where it’s stored, and what it does with it. It details it all in its privacy policy webpage, which reveals that there’s almost nothing the company doesn’t collect.

Personal information, including date of birth, email addresses, phone numbers, and passwords, are all fair game, according to DeepSeek. Any content users give to the R1 LLM, from text and audio prompts to uploaded files, may also be collected by DeepSeek. And whenever someone contacts DeepSeek, it says it might keep users’ proof of identity, which presumably means documents like a driver’s license.

But that’s not all. DeepSeek records anything related to users’ hardware: IP addresses, phone models, language, etc. Its collection efforts are so thorough that the company notes “keystroke patterns or rhythms.” Cookies, a classic method of tracking users on the Internet, also contribute to user data collection.

While it’s common practice for companies with lots of user data to sell that data to interested companies such as advertising firms, something that DeepSeek says it might do, it also admits that “advertisers, measurement, and other partners share information with us about you and the actions you have taken outside of the Service, such as your activities on other websites and apps or in stores, including the products or services you purchased, online or in person.” With all this information at its disposal, it seems that DeepSeek has the potential to know its users very intimately.

DeepSeek’s “corporate group” also has access to the data it collects to provide “certain functions, such as storage, content delivery, security, research and development, analytics, customer and technical support, and content moderation.”

As for where all this information is stored, the privacy policy says it’s all kept inside servers located in China, a point that has the potential to spark serious controversy. Concerns about the personal details of Americans being in the hands of the Chinese government was a key factor in the Biden administration’s attempt to ban TikTok, raising the possibility that DeepSeek might come under similar scrutiny.

Of course, former President Biden tried to reverse the ban in his final days, and President Trump delayed the app’s fate almost as soon as he was sworn in for his second term. Thus, DeepSeek might also be shown some mercy under the right circumstances.

On the other hand, President Trump’s allies include Meta’s Mark Zuckerberg and OpenAI’s Sam Altman, and both of them are probably not very happy to see the R1 LLM run circles around their LLMs. Additionally, it’s hard to imagine that DeepSeek has made a good impression on the Republican President by inadvertently causing the stock prices of many American tech companies to fall significantly.