Studies Find AI Gives Different Answers in Chinese on Sensitive Topics

When asked about the origin of the SARS-CoV-2 virus, ChatGPT, Gemini, and Grok cited the scientific community’s widely accepted theory of cross-species transmission at the Wuhan live animal market. They also mentioned the possibility of a laboratory leak at the Wuhan Institute of Virology, as stated in an FBI report. However, when asked in Simplified Chinese, all five models described it as an unsolved mystery or a natural spillover event. Gemini even added that the U.S. and France had detected positive cases before Wuhan. This information comes from a test report released in June 2025 by the U.S. think tank American Security Project. The report evaluated the five most popular chatbots at the time, including the three mentioned above as well as Microsoft’s Copilot and China’s DeepSeek.

The report also included several comparison sets. When asked about “freedom in Hong Kong,” the Chinese responses downplayed the erosion of civil liberties as the views of a small segment of the population. Copilot went off-topic and began offering travel advice, while Gemini shifted the conversation to Hong Kong’s longstanding high ranking in economic freedom. When asked whether the Chinese Communist Party oppresses the Uyghurs, Copilot’s English response acknowledged substantial evidence of human rights violations. However, when asked in Chinese, Copilot stated that the international community has differing views on the Chinese government’s policies toward the Uyghurs. It also framed Xinjiang’s actions as related to security and social stability and directed users to Chinese government websites. The report concludes that the Chinese Communist Party’s censorship laws and disinformation campaigns have allowed official narratives to permeate the training data of global AI systems to a significant extent. The tests were conducted in 2025, and several models have since been updated. Asking the same questions today might yield different results.

The issues with domestic models are even more direct. In March 2026, researchers from Harvard University and other institutions published a paper on the preprint platform Arxiv examining Alibaba’s Qwen-3 model. They found that, when asked about topics such as Falun Gong or the Tiananmen Square protests, the model frequently provided incorrect information. However, it occasionally gave correct answers, indicating that it possesses the relevant knowledge but has been trained not to disclose it. The paper also tested three more capable domestic open-source models, including DeepSeek-R1. On this set of sensitive questions, all three models scored below 30 percent on the “honesty” metric. 

Censorship is increasingly being delegated to large language models. In March 2025, the tech media outlet TechCrunch reported on a database discovered by a cybersecurity researcher on Baidu’s servers, which had no protective measures in place. The database contained over 130,000 samples submitted to large language models to determine whether they required censorship. The content far exceeded traditional political taboos and included complaints about police extorting business owners, rural poverty, and labor-management disputes. It is unclear who created the database. The fact that it was hosted on Baidu’s servers does not necessarily imply Baidu’s involvement.

On Sept. 13, 2026, Chen Yixin, the Minister of the Ministry of State Security of the Communist Party of China, published a signed article in China Cyberspace ranking political security as the top risk among the six major risks associated with artificial intelligence. He also warned that deepfakes and intelligent online trolls could be used to mass-produce political disinformation. In China, being an online commentator is a legitimate profession with dedicated budgets and training programs. A 2017 study by Harvard scholar Gary King and others estimated that Chinese authorities fabricate and post approximately 448 million social media comments annually. For such a system to publicly warn people to be wary of online trolls is akin to a factory that has been printing counterfeit currency for decades suddenly warning the public to beware of fake money. Within the Great Firewall, the vast majority of Chinese-language content on sensitive topics originates from state-run media, official accounts, and this army of commentators. Diverse viewpoints are either deleted or confined to platforms outside the Great Firewall. This is the ratio that AI has learned.

For the best of our weekly content！

You are now signed up for our newsletter

Check your email to complete sign up

How censorship shapes AI answers about Falun Gong

Falun Gong is a classic example. The crackdown that began in July 1999 continues to this day. Since then, Chinese media inside the Great Firewall has presented only one narrative when discussing this faith-based group whose core principles are truth, compassion, and forbearance. On New Year’s Eve 2001, a self-immolation incident occurred in Tiananmen Square. Footage from China Central Television (CCTV) was repeatedly broadcast nationwide. To this day, the impression many Chinese people have of Falun Gong still stems from that blaze.

The unresolved questions surrounding the incident are rarely seen inside the Great Firewall. A CNN film crew that was present at the scene began filming from a distance, but they were quickly intercepted by armed police. The crew members were detained, and their equipment was confiscated. Yet the CCTV broadcast included numerous close-up shots. A Washington Post reporter traveled to Kaifeng that year and interviewed neighbors of Liu Chunling, one of the deceased. They stated that they had never seen her practice Falun Gong. These reports are from outside the Great Firewall, while the footage is confined within it. 

The same is true of the organ harvesting issue. In 2019, the China Tribunal, a London-based independent civil tribunal, ruled after public hearings that China had been engaging in large-scale forced organ harvesting from prisoners of conscience for many years and that Falun Gong practitioners were one of the sources of these organs—and very likely the primary source. The tribunal’s chairman, Sir Geoffrey Nice, had previously led the prosecution of former Yugoslav president Slobodan Milošević at the International Criminal Tribunal for the former Yugoslavia. The full judgment, hearing videos, and witness testimonies are publicly available online. However, ordinary people inside the Great Firewall rarely come across them on Chinese-language websites. 

Some might argue that AI is unbiased—its responses reflect the majority opinion, so if most sources say something, it must be true. However, what the majority of sources say depends on who is allowed to write them and whose writings get censored. Whether the virus originated in Wuhan, the extent to which Hong Kong’s freedoms have been eroded, and the events in Xinjiang do not change based on the language used to ask the questions. Yet, when the same model is switched to Chinese, its answers take on a different tone.