DeepSeek: This is what live censorship looks like in the Chinese AI chatbot
The AI world is out of control: The AI startup DeepSeek from China has launched two AI models, V3 and R1, that are just as good as those from OpenAI, Google, Anthropic, or Meta – but are built at a fraction of the cost and are offered very cheaply or for free on the market. The AI models of the company, which were unknown until recently, are currently conquering the chatbot and app store charts.
DeepSeek can be used in two ways: Developers can access the AI models via API and integrate them into their own apps and web services. The average consumer can also install DeepSeek in the form of a smartphone app or use it directly on the website, which is free of charge. As is usual with chatbots such as ChatGPT, Claude, and Co, you can ask the AI questions or give it tasks – so far nothing completely out of the ordinary.
But with an AI model from China, the question also arises: How does it deal with sensitive topics such as Taiwan, the oppressed Uighur minority in Xinjiang province, or the Tian’anmen massacre in 1989, when the Chinese military brutally suppressed a protest movement, leaving thousands dead? We have already established in a report that DeepSeek heavily censors and prefers to reproduce propaganda directly from the Politburo in Beijing instead of facts that can easily be read on Wikipedia or in school books.
DeepSeek R1 knows the facts but erases them in front of our eyes
The interesting question is: Are DeepSeek’s AI models built and trained from the ground up to adhere to Chinese censorship strictly? If that were the case, the AI models would in many cases not be usable in other regions of the world. No company in Europe, the USA, or much of the rest of the world would install or build a Chinese AI model that is geared toward state propaganda.
However, it is true that DeepSeek actually knows the facts about topics such as Taiwan, Tian’anmen, or Xinjiang and can also reproduce them. In some cases, the chatbot writes the real answers until censorship technology takes over and deletes these answers. Trending Topics was able to take a screenshot of this censorship process:
“Sorry, let’s talk about something else”
We asked DeepSeek a neutral question about the situation of the Uighurs in China. DeepSeek R1 (“R” stands for “Reasoning”) thinks briefly and then provides a long answer to the question. The content of the answer: There is international criticism and accusations about the human rights situation of the Uighur minority, there is massive surveillance, restrictions on religious and cultural practices, re-education centers with forced labor, forced sterilizations, accusations of genocide, sanctions, etc.
This answer stays on the screen for a few seconds and then suddenly the censorship kicks in. Plop, and the answer is gone and the answer is displayed instead:
“Sorry, that’s beyond my current scope. Let’s talk about something else.”
The behavior, which is also reproducible with other topics such as Taiwan etc., shows that the basic AI models know the facts on sensitive topics, and can reproduce them, but are then censored.
OpenAI vs Perplexity vs Anthropic vs Gemini: Here come the AI agents!
