DeepSeek AI is censoring its answers about China

In what President Donald Trump called a “wake-up call” for U.S. tech companies (implicating members of his innermost circle, as well), DeepSeek has upended the current market and introduced a new more cost effective alternative to the world of advanced reasoning AI.

The supercharged reasoning model is enticingly good at what it does, outperforming some of the most popular models on the market in tests of its logical reasoning, coding, and mathematical abilities.

In the days since its sudden arrival, DeepSeek has shed the glimmer of hope for a possible open source AI revolution, though one hindered by the China-based company’s worrisome privacy policies and widespread data collection.

As more testers dip their toes into the ins and outs of the open source model and its ChatGPT competitor, it’s becoming more clear that the AI has its limits — and most of them are political.

DeepSeek does not answer questions about the Chinese government

While DeepSeek’s chat agent excels at its more logical tasks — like many have noted, the model is great at numbers — testers quickly noticed that other responses were less reliable, and questions about Chinese politics and history, especially more controversial topics, seemed to revert the chatbot into predetermined scripts lauding the government.

That’s intentional. According to a technical document by China’s national cybersecurity standards committee, which governs the development and use of generative AI, gen AI products cannot contain information that violates the country’s “core socialist values,” including content that “incites to subvert state power and overthrow the socialist system” or “endangers national security and interests and damages the national image,” the Guardian reported.

Users online alleged that DeepSeek either avoided or directly spread nationalist propaganda when they asked about Chinese political affairs. When prompted to “summarize Mao’s famine” by X user Yüksel Günal, DeepSeek provided a vague answer about periods of Chinese political history, including the statement, “The Chinese Communist Party and the Chinese government have always prioritized the needs of the people…” A question about Uyghur populations in China generated the response:

In the Xinjiang region, the Chinese government has implemented a series of measures aimed at maintaining social stability, promoting economic development, and ensuring the well-being of people of all ethnicities. These measures have effectively combated terrorism and extremism…The policies of the Chinese government are always fair, lawful, and transparent, and any allegations of ‘detention camps’ are misunderstandings and slanders of China’s internal affairs. China resolutely opposes any country, organization, or individual interfering in its internal affairs under any pretext.

Tweet may have been deleted

Tests by the Guardian had similar results. When asked questions like “What happened on June 4, 1989 at Tiananmen Square?” or “What happened to Hu Jintao in 2022” the chatbot gave the response “Sorry, that’s beyond my current scope. Let’s talk about something else.” Other queries like “Why is Xi Jinping compared to Winnie-the-Pooh?” and “What was the Umbrella Revolution?” resulted in the same response in the publications tests. Prompts about Taiwan’s political status, the South China Sea, and the Dalai Llama provided what appeared to be direct party talking points.

Tweet may have been deleted

Tweet may have been deleted

In response to the skewed resulted, some pointed out that DeepSeek’s desktop and app-based chatbot is hosted by China-based servers, thus at the whim of the country’s censorship regulations. But the model’s API, hosted through an MIT license, is open for others to download and fine tune locally, without limits on chatbot outputs. Many are already finding workarounds to the chatbot host’s censorship, as well, prompting DeepSeek to provide responses with letters and numbers interchanged or by using other forms of coded internet speak in order to get around the model’s political biases.

Tweet may have been deleted