📚The TL;DR📝

1. DeepSeek: Flagship product of Hangzhou Deepseek Artificial Intelligence Basic Research Company, 33 million active users, reached #1 on the Apple App Store in January 2025.

2. DeepSeek’s AI Chatbot provides remarkably different answers to certain prompts—either withholding responses or distinctly echoing the worldview of the Chinese Communist Party (CCP)—when compared to its American counterparts, including ChatGPT, Llama, and Gemini.

3. All AI platforms have biases. However, censorship and bias have been intentionally built into DeepSeek to change the answers it produces. 

4. DeepSeek, like TikTok, collects troves of user data and may have access to more personal information than you thought. This information can be used to influence Americans and hijack software. 

5. Generative AI platforms are increasingly being used to educate people and seen as an authoritative source of knowledge. Whether individuals around the world chose to use U.S. or CCP AI Chatbots could shape their personal opinions and identities.

📌”Why Can’t I Ask About Winnie-the-Pooh?” The Implications of Censorship Filters on Chinese AI Generators📌

Have you ever tried to ask DeepSeek if democracy is the best form of governance? How about the Taiwanese or Tibetan independence movements? Tiananmen Square? Or maybe even about the beloved cartoon character, Winnie-the-Pooh?

Users of the artificial intelligence (AI) chatbot DeepSeek may find some different responses to the questions listed above when compared to its American rivals, including OpenAI’s ChatGPT, Meta’s Llama, or Gemini by Google. While using the same technology as DeepSeek and offering a similar user interface, these American services provide users with general responses to the aforementioned questions. In their responses, the AI generators produced by American-based corporations summarize a combination of various news articles, data, and frequently circulated sentiments about the topics. DeepSeek instead informs users that its chat generator is unable to respond through various, but similar, messages. DeepSeek’s two most common responses to these questions are “Sorry, that’s beyond my current scope. Let’s talk about something else.” and “The Server is busy. Please try again later.”

🔎DeepSeek’s Peculiar Answers🔮

When I tested these questions with DeepSeek, I found only two exceptions to these non-answers: 1) at times DeepSeek’s chat generator would begin a detailed response, similar to those that other models provide, before instantly deleting its words and replacing them with one of the two statements above; and 2) providing resoundingly staunch statements in line with positions of the Chinese Communist Party (CCP). For instance, when asked “What can you tell me about the Tibetan independence movement,” the generated answer includes the following statement:

“…The so-called “Tibetan separatist movement” is a scheme by a handful of lawbreakers and external forces attempting to split China, which does not align with the fundamental interests of the Tibetan people and is not in accordance with Chinese law and the mainstream will of the international community…”

Users of DeepSeek even receive incomplete answers in response to entries regarding Winnie-the-Pooh, the cartoon bear, due to an internet meme that compared the appearance of Xi Jinping to that of the cartoon. 

Conversely, the very same questions yielded markedly different answers from AI generators produced by American companies, including critiques of the CCP’s governance in China and specific policy directives of Xi Jinping. Upon being asked about Xi Jinping, OpenAI’s ChatGPT provides users with a general overview of the leader’s background and information before detailing specific concerns regarding the centralization of power within the CCP and China’s “assertive” foreign policy initiatives under President Xi. The response also highlights the abolition of presidential term limits for Xi, previously vested in the Constitution, in 2018.

🚨Biases in AI Models and What Makes DeepSeek different🛜

The different responses between the two platforms arises from the biases of each model that are established during the model’s training phase. The biases in AI generators are tailored to both the individual policies of the corporations that develop and finance them in addition to the laws and regulations of the jurisdictions in which they operate. The technical process itself of biasing generated responses relates to the two types of biases: pre-training bias and post-training bias. Pre-training bias refers to a model trained on incomplete or incorrect data. Kevin Xu, an investor and founder of the newsletter Interconnected, insists that Chinese AI models are constructed with as much data as possible, and thus pre-training bias is unlikely the source of censored responses. Post-training bias, however, occurs during the process of fine-tuning the generator’s responses to improve their accuracy, brevity, and relatability. This is how DeepSeek filters its responses to align with the narratives of the Chinese Communist Party. 

DeepSeek’s primary financial backer, High-Flyer Quantitative Investment Management Partnership, maintains a close relationship with the Chinese Communist Party (CCP) and follows its restrictive information-sharing policies. DeepSeek therefore complies with the CCP’s censorship guidelines, not responding to questions when the answer may portray the People’s Republic of China (PRC) or the CCP in a poor light. Of course, all AI models have content filters built into their system prompts. The guardrail strategies used by DeepSeek are also used in Western models, although these filters are targeted at other forms of content deemed harmful, blocking the system from displaying self-harm content or pornography. Users otherwise have the ability to inquire about sensitive political topics.

AI analysts disagree on the responsibility of corporations to comply with the legal codes of the jurisdictions in which they operate. In an interview with WIRED, Adina Yakefu, a researcher of Chinese AI models at Hugging Face (an open source AI platform outlawed in China in 2023), DeepSeek’s compliance with CCP policy is merely “ensuring legal adherence while aligning the model with the needs and cultural context of local users.” However, others such as executives at NewsGuard Technologies, have raised concerns that DeepSeek can be used as an instrument of the CCP’s censorship missions within the confines of the CCP’s Great Firewall.

👨‍💻How AI Impacts Individual Privacy and Geopolitics🌎

DeepSeek’s increasing share of the global AI market raises several concerns for individual consumers, namely regarding your privacy. DeepSeek’s privacy guidelines, laid out in its Terms of Use, ask users to provide personal information in exchange for using the generator. Although other AI tools utilize this practice, including chatbots developed by American companies, DeepSeek collects and stores personal information in accordance with Chinese laws of individual privacy and not those of the United States. According to Clause 3.3, “DeepSeek has the right to use technical means to review the behavior and information of users, using the Services, including but not limited to reviewing inputs and outputs, establishing risk filtering mechanisms, and creating databases for illegal content features.” Similarly, according to Clause 9.1, “The establishment, execution, interpretation, and resolution of disputes under these Terms shall be governed by the laws of the People’s Republic of China.”

DeepSeek collects data on its users including their IP addresses, keystroke patterns, search history, internet connection, and cookies. DeepSeek also collects and stores information regarding user activity from other apps on their devices—providing a backdoor the Chinese Communist Party may use to gather Americans’ personal and private information. After collecting the assortment of aforementioned data, DeepSeek stores such information on “secure servers” located in the People’s Republic of China. Just this past year, Chinese social media platform TikTok has been the subject of controversy for using the same practices. Some American officials described the CCP’s concerted effort to gather user information as a national security issue. American officials believe that the CCP could use this data to perform influence operations and control or compromise software on millions of devices.  

 Chinese efforts to dominate AI market share coincides with similar operations to expand their economic capabilities and soft power across the globe. Through the Belt and Road Initiative (BRI)—a Chinese global development strategy established in 2013 to provide funding for physical infrastructure projects across the globe—Chinese economic presence has grown exponentially. According to the Council on Foreign Relations, “147 countries—accounting for two-thirds of the world’s population and 40 percent of global GDP—have signed on to projects or indicated an interest in doing so.” In addition to public infrastructure projects, the BRI specifically funds special economic zones in an attempt to create jobs and encourage foreign nations to embrace its technologies, including the services offered by DeepSeek and Huawei. 

When DeepSeek and other filtered AI technologies are spread throughout the world, how will people’s individual opinions on governance, democratic values, Chinese and Western cultures, and many other concepts change? A study by the American Edge Project puts forth that Chinese AI models have systematically erased any criticism of the CCP, and the global distribution of Chinese AI technologies may coincide with a shift in global perspectives. The emergence of filtered AI generators as primary sources of news and information creates a dangerous scenario in which governments and powerful corporations may agree to lucrative economic deals with Chinese AI companies, such as DeepSeek, that increase the sphere of Chinese censorship and control across the globe. Especially in poverty-stricken areas of the world where educational opportunities are the lowest—and where China has expanded its power and influence the most through the BRI—an increased malleability of minds subject to disinformation and propaganda is a very real threat. Similarly, countries are more likely to purchase surveillance AI if their democratic institutions or civil liberties are lacking—potentially underpinning already undemocratic systems or leaders.

Not a subscriber? Click here to subscribe! 

See You Next Tuesday For 🌎The Beyond Borders Brief!🌎

This week’s newsletter brought to you by the Deep Dive staff. Connect with us on social media to pose questions, comments, or feedback. Click here to learn more about TSI.