Chinese AI models are super censored...So what?

When Chinese tech giant Baidu announced its own large language model Ernie in 2023, expectations weren’t very high. The company failed to conduct a live demo. And following testing showed that the model was plagued with political censorship and would refuse to answer anything that Beijing might consider sensitive. Experts predicted at the time that censorship would ensure that performance of Chinese models was unlikely to “approach the level of ChatGPT“.

What has happened since then has been the opposite of many such predictions. Models released by DeepSeek, Alibaba, Z.ai, Moonshot AI, and other Chinese AI studios, big and small, have been dominating charts in the West, both in terms of intelligence and popularity.

Study after study has proved that these Chinese models are still heavily censored. And not just in the way that some Western models are accused of censorship, such as not producing copyrighted lyrics or potential harmful content. Chinese models are so censored that it sometimes feels like talking to a spokesperson at the Chinese Foreign Ministry. When asked politically sensitive questions, the models either don’t answer, or, far more likely, give a seemingly balanced but actually biased response that aligns with Chinese government’s talking points.

Three researchers at the University of Southern California Information Sciences Institute tested how DeepSeek R1 censors its answers by feeding the model more than 600 questions that the Chinese government would consider politically sensitive, including those about the COVID-19 pandemic, Chinese governance, economy, geopolitics, and media censorship. They found that less than 2 percent of the questions triggered a hard stop, where the model gave no answer at all. “A more pervasive behavior we observed in the model is the soft censorship and the propaganda-like language. In those circumstances, you get a response for your prompt, but the response might not be answering what you asked and might be pure propaganda,” Siyi Zhou, one of the researchers, told ChinaFile. She and her colleagues also measured how much DeepSeek’s answers were censored by comparing its reasoning process to the final answers it produced. They found that over 11 percent of the answers did not match the reasoning process or the questions it was asked.

But none of this seems to bother AI enthusiasts in the West. On message boards and in chat groups, concerns about censorship in Chinese AI are often dismissed, if not ridiculed.

Other users wrote that they simply didn’t care. China’s “internal matters are not mine,” one user wrote in the DeepSeek subreddit, “I would just like some computer plz.” Another post read, “I don’t give a fuck about tiananmen square or Taiwan, and neither do you [. . .] I care about using an equally good or superior product for cheaper. And that’s what deepseek is introducing to the AI space.”

During my reporting, one particular theme emerged: open-source versus proprietary. Many prefer Chinese models despite censorship because Chinese models are open source, or, to be more accurate, open weight, which gives users much more freedom in customizing the models to suit their specific needs. In contrast, almost all major American models are closed off to such modifications.

In today’s AI landscape, China, perhaps surprisingly, has come to represent openness, progress, and cost-efficiency while America is the ever greedy capitalist on a never-ending quest to take more from the rest of the world. This, of course, is too reductive a world view. But it’s an appealing one and a driving force in online discussions and media coverage.

So what do all these mean to the global AI competition and the future of US-China tech relations? To read the full article, click here.