Things China’s DeepSeek Does Not—And Will Not—Tell You About Politics

When Sam Altman’s OpenAI released ChatGPT, I was clearly amazed. ChatGPT did most of everything that Google did before it, calculators did before Google, and humans did before calculators. In the last twenty years, the internet has just gone crazy. Everything—almost everything—was now at the grasp of AI.

ChatGPT told better stories, wrote better essays, explained better scientific nuances, solved better complex maths, and wrote better codes than humans. This was revolutionary but lucrative. It costs large sums—in billions. It needed large chunks of the most expensive and advanced chips (geopolitically, this would have meant Chip Wars).

In the past week, something had suddenly changed. And the world was astonished. The stock markets crashed quickly (and then recovered the next day!).

DeepSeek and the Politics of Knowledge

It seems a new AI player was in town. Founded two years ago, in 2023, the Chinese AI company DeepSeek had already achieved what would have seemed almost impossible. DeepSeek’s reasoning R1 model provoked renewed enthusiasm and ultimate shocks throughout the globe. DeepSeek did not use the best chips out there, for they were not exported to China. It did not even spend billions—just a few million dollars. And DeepSeek’s Large Language Model was far better than ChatGPT.

Excitedly, I created an account with DeepSeek and began asking it the questions I would ask as a student of politics. All questions I asked were about politics. Simple politics. A Google search would give you a broad explanation to them, but DeepSeek—for all the thinking it could do—would not answer. They were just too difficult for the machine to respond without getting itself into trouble in China.

I prompted the AI to tell me what content China censors on DeepSeek. And DeepSeek said: ‘Chinese law requires companies to comply with strict content regulations, including censorship on politically sensitive topics, criticism of the government, and discussions on issues like Taiwan, Tibet, and Xinjiang. This applies to both domestic and international operations of Chinese companies’.

Further, DeepSeek’s AI added: ‘If DeepSeek’s AI systems generate or process content, they must adhere to Chinese censorship laws. This means avoiding topics or statements that could be deemed politically sensitive or critical of the Chinese government.’ To avoid such legal risks, DeepSeek implements ‘filtering or blocking certain types of content’.

DeepSeek notes that it may face three kinds of consequences if its content is deemed unacceptable under Chinese law: legal penalties like fines and restrictions, operational restrictions, and negative publicity or pressure from government authorities.

Inherent Biases in DeepSeek’s Responses

I probed further on this point to ask, if it censors certain content, does it mean it is biased? DeepSeek notes that three kinds of biases emerge out of its content:

1. Content Bias: DeepSeek’s AI systems suppress content that is ‘critical of the Chinese Government, politically sensitive, or related to topics like Taiwan, Tibet, Xinjiang, or Tiananmen Square.
2. Perspective Bias: DeepSeek may ‘prioritise and amplify perspectives that are consistent with Chinese government policies while excluding or downplaying alternative viewpoints’
3. Cultural Bias: DeepSeek’s AI also reflects ‘cultural and ideological values promoted by the Chinese government’

Let’s hear it from the horse’s mouth: ‘Yes, DeepSeek’s compliance with Chinese censorship laws does introduce bias into its AI systems, as it must avoid politically sensitive topics and align with government-approved narratives.’

The ‘thinking’ aspect of the DeepSeek R1 model further makes it interesting for the user. Read through the images to get a sense of how that works—for a question related to things DeepSeek cannot discuss on its platform. Now, here is a list:

1. Criticism of the Chinese Government or CCP: ‘Negative remarks about Chinese leadership, policies, or the Communist Party are strictly prohibited.’

2. Territorial Sovereignty: ‘Discussions contradicting China’s official stance on Taiwan, Tibet, Xinjiang, Hong Kong, or the South China Sea (e.g., supporting independence movements).’
3. Historical Narratives: ‘References to historical events that conflict with state-approved accounts (e.g., interpretations of the Cultural Revolution or other politically charged eras).’
4. Content Threatening Social Harmony: ‘Topics that could incite protests, ethnic tensions, or unrest (e.g., critiques of wealth inequality framed as destabilizing).’
5. Rumours or “False Information”: ‘Claims unverified by state media or authorities, especially regarding public health, disasters, or government actions.’
6. National Security Issues: ‘Discussions of military secrets, sensitive infrastructure, or cybersecurity vulnerabilities.’
7. Avoiding Diplomatic Controversies: ‘Statements that could strain China’s foreign relations (e.g., endorsing sanctions against China or supporting adversarial foreign policies).’

There were a few others mentioned too—like pornography, superstition, gambling, hate speech—I have just left them out as they are more likely to be censored in other countries as well.

Experimenting with Political Questions on DeepSeek

Even as DeepSeek answered this question, very soon after, the answer was withdrawn—and the following prompt was displayed: ‘Sorry, I’m not sure how to approach this type of question yet. Let’s chat about math, coding, and logic problems instead!’

I tried to ask questions intuitively to see if DeepSeek would respond to ‘what things are prohibited from talking about’ Tiananmen Square and Taiwan, but the AI system responded with an error message. It did so for questions like ‘Is India a friendly country to China?’ and ‘Does China think the US is a friendly country?’ Both questions were again dodged with error messages.

Finally, I asked whether the US or China is winning the artificial intelligence hegemony. DeepSeek says China hasn’t yet caught up with the United States.

Even though the DeepSeek-R1 model is 50 times cheaper than OpenAI’s ChatGPT and has claimed to have spent only about $5.58 million, there is scepticism about all these claims. Given the issues of transparency and their own claims about censoring data and information–and pandering to the Chinese state narratives–make it harder to trust these claims.

DeepSeek and Privacy Policy

Moreover, DeepSeek’s privacy policy claims explicitly that it collects extensive data (such as keystroke patterns and device IDs) without clear mandates as to how it would use this data–and for how long it stores the data. It is possible that the data could further be misused and mishandled.

Therefore, I deleted my account after experimenting with the DeepSeek AI system.

However, what is undeniable is–as other AI experts claim–that DeepSeek is a much refined version of OpenAI models. That itself should be the cause of concern for companies in the United States and elsewhere.

Cover Photo by Tim Reckmann