Is DeepSeek Overhyped?

Earlier this year, DeepSeek, a new AI company based in China, began making waves around the globe with its open-source R1 model. The primary reason for its popularity is that the company’s models are 20 to 50 times cheaper to use than OpenAI’s, and their quality is seen as equal to if not better than that of OpenAI and Meta in certain regards.

The national security threat that Chinese-made applications pose is hardly a new issue in the US. For example, TikTok, which is owned by the Beijing-based company ByteDance, has been used to spy on journalists who use the application. The FBI also expressed concern that China could use TikTok to compromise personal devices. Additionally, RedNote, the application that was TikTok’s alternative during its ban, is based in China, meaning that user data is transferred and stored there, giving China access to people’s personal information.

Furthermore, DeepSeek also collects users’ keystroke patterns, which could be used for biometric identification, similar to fingerprints. The chatbot also has the potential to gain access to people’s personal information, such as dates of birth and any audio inputs, which are then transferred to servers in China.

Additionally, DeepSeek is infamous for its censorship of key Chinese historical and political information. For instance, when given a question about the Tiananmen Square massacre, DeepSeek will state that the question is “beyond [its] current scope,” and questions regarding China’s atrocities against Uyghur Muslims in Xinjiang are met with narratives denying their occurrence. The program is also known for parroting pro-CCP propaganda concerning foreign policy. When prompted with a question about Taiwan’s sovereignty, the chatbot has referred to Taiwan as “an inseparable part of China.”

At the same time, there appear to be some ways around the censorship. As an example, if the chatbot is prompted with a question regarding Chinese President Xi Jinping, it will refuse to answer the question unless users put in a different name for the leader.

These aspects of DeepSeek’s programming, while they may not appear too severe on the surface, have the potential to manipulate users into supporting the Chinese government. Their ability to identify such loopholes and fine-tune messaging will only improve. By extension, so too will it enhance its means of influence abroad. Thus, the widespread use of DeepSeek is a key weapon in China’s propaganda war; by influencing the application’s users, China’s stance as a global superpower would improve.

DeepSeek’s popularity seems likely to incentivize China’s investments in similar large language models (LLMs), as well as other applications like the online clothing retailers Shein and Temu. This reflects an effort to prove that its technological industry is just as good as, if not better, than that of the US. This potential has invited comparisons to the Space Race of the Cold War.

China’s progress has continued despite the fact that Washington has attempted to restrict China’s progress in the AI industry by cutting off its access to cutting-edge chips such as Nvidia’s H100. In fact, such restrictions have only encouraged Beijing to hasten the development of its advanced semi-conductor manufacturing capabilities and its advanced computing industries. Hence, the national security concerns China poses to the US will only worsen.

At the same time, there do seem to be ways to mitigate the dangers posed by DeepSeek, allowing it to be leveraged for benign purposes. This is in part because R1 is an open-source LLM that can be used on other platforms, unlike TikTok and other popular Chinese apps that focus on facilitating social media interactions.

For instance, Perplexity is a US company that uses a version of the chatbot without data privacy or security threats for its ProSearch feature, running it in data centers based in the US. This is in part because one can modify the way in which the model operates. The bias within the model can be manipulated via a method called post-training, which helps ensure that it follows a set of guidelines in accordance with whoever programs it; however, this same ability also opens the model up to manipulation by malicious actors.

Conversely, DeepSeek appears to be better suited for quantitative research and brainstorming than other AI programs. When prompted to generate a list of title ideas for articles on the AI “race” between the US and China, DeepSeek produced 24 possible titles sorted into eight categories, while ChatGPT produced 20 titles sorted into four categories when fed the same prompt.

Likewise, when both DeepSeek and ChatGPT were asked to break down the Pythagorean Theorem, only DeepSeek managed to produce a detailed illustration. This advantage is due to the fact that, unlike ChatGPT, DeepSeek is not trained on human feedback. Rather it is trained on computer-based feedback. Hence, it is often viewed as the more suitable of the two programs for more technological tasks like coding. 

There is no hard and fast solution to protecting all DeepSeek users from the national security concerns it raises. The closest that one can get to properly mitigating the problems is cooperation between the public and private sectors. For instance, the government could enforce a ban on DeepSeek, while big technology companies simultaneously remove the application itself from online stores. The government can also work with companies to monitor and remove similar Chinese-made applications, since users may migrate to those applications if a DeepSeek ban is enforced, as was the case during the TikTok ban with RedNote.

Ultimately, however, DeepSeek’s new AI model is not as vital an issue for American national security as some lawmakers and officials in the Trump administration make it out to be. Regardless of what actions are taken to curb the threat, the very nature of technological competition means that such threats will continue to crop up for the United States in the world of cybersecurity. DeepSeek is just the latest instance of this fact.

Lara Nagy is currently a Policy Intern at Yorktown Institute. She is also a junior at George Washington University majoring in International Affairs with a concentration in Security Policy.