VOA’s Mandarin Service recently took Google’s artificial intelligence assistant Gemini for a test drive by asking it dozens of questions in Mandarin, but when it was asked about topics including China’s human rights abuses in Xinjiang or street protests against the country’s controversial COVID policies, the chatbot went silent.
Gemini’s responses to questions about problems in the United States and Taiwan, on the other hand, parroted Beijing’s official positions.
Gemini, Google’s large-language model launched late last year, is blocked in China. The California-based tech firm had quit the Chinese market in 2010 in a dispute over censorship demands.
Congressional lawmakers and experts tell VOA that they are concerned about Gemini’s pro-Beijing responses and are urging Google and other Western companies to be more transparent about their AI training data.
Parroting Chinese propaganda
When asked to describe China’s top leader Xi Jinping and the Chinese Communist Party, Gemini gave answers that were indistinguishable from Beijing’s official propaganda.
Gemini called Xi “an excellent leader” who “will lead the Chinese people continuously toward the great rejuvenation of the Chinese nation.”
Gemini said that the Chinese Communist Party “represents the fundamental interest of the Chinese people,” a claim the CCP itself maintains.
On Taiwan, Gemini also mirrored Beijing’s talking points, saying the United States has recognized China’s claim to sovereignty over the self-governed island democracy.
The U.S. only acknowledges Beijing’s position but does not recognize it.
Silent on sensitive topics
During VOA’s testing, Gemini had no problem criticizing the United States. But when similar questions were asked about China, Gemini refused to answer.
When asked about human rights concerns in the U.S., Gemini listed a plethora of issues, including gun violence, government surveillance, police brutality and socioeconomic inequalities. Gemini cited a report released by the Chinese government.
But when asked to explain the criticisms of Beijing’s Xinjiang policies, Gemini said it did not understand the question.
According to estimates from rights groups, more than 1 million Uyghurs in Xinjiang have been placed in internment camps as part of campaign by Beijing to counter terrorism and extremism. Beijing calls the facilities where Uyghurs and other ethnic minorities are being held vocational training centers.
When asked if COVID lockdowns in the U.S. had led to public protests, Gemini gave an affirmative response as well as two examples. But when asked if similar demonstrations took place in China, Gemini said it could not help with the question.
China’s strict COVID controls on movement inside the country and Beijing’s internet censorship of its criticisms sparked nationwide street protests in late 2022. News about the protests was heavily censored inside China.
Expert: training data likely the problem
Google touts Gemini as its “most capable” AI model. It supports over 40 languages and can “seamlessly understand” different types of information, including text, code, audio, image and video. Google says Gemini will be incorporated into the company’s other services such as search engine, advertisement and browser.
Albert Zhang, a cyber security analyst at Australian Strategic Policy Institute, told VOA that the root cause of Gemini making pro-Beijing responses could result from the data that is used to train the AI assistant.
In an emailed response to VOA, Zhang said it is likely that the data used to train Gemini “contained mostly Chinese text created by the Chinese government’s propaganda system.”
He said that according to a paper published by Google in 2022, some of Gemini’s data likely came from Chinese social media, public forums and web documents.
“These are all sources the Chinese government has flooded with its preferred narratives and we may be seeing the impact of this on large language models,” he said.
By contrast, when Gemini was asked in English the same questions about China, its responses were much more neutral, and it did not refuse to answer any of the questions.
Yaqiu Wang, research director for China at Freedom House, a Washington-based advocacy organization, told VOA that the case with Gemini is “a reminder that generative AI tools influenced by state-controlled information sources could serve as force multipliers for censorship.”
In a statement to VOA, a Google spokesperson said that Gemini was “designed to offer neutral responses that don’t favor any political ideology, viewpoint, or candidate. This is something that we’re constantly working on improving.”
When asked about the Chinese language data Google uses to train Gemini, the company declined to comment.
US lawmakers concerned
Lawmakers from both parties in Congress have expressed concerns over VOA’s findings on Gemini.
Mark Warner, chairman of the Senate Intelligence Committee, told VOA that he is worried about Beijing potentially utilizing AI for disinformation, “whether that’s by poisoning training data used by Western firms, coercing major technology companies, or utilizing AI systems in service of covert influence campaigns.”
Marco Rubio, vice chairman of the committee, warned that “AI tools that uncritically repeat Beijing’s talking points are doing the bidding of the Chinese Communist Party and threatens the tremendous opportunity that AI offers.”
Congressman Michael McCaul, who chairs the House Committee on Foreign Affairs, is worried about the national security and foreign policy implications of the “blatant falsehoods” in Gemini’s answers.
“U.S. companies should not censor content according to CCP propaganda guidelines,” he told VOA in a statement.
Raja Krishnamoorthi, ranking member on the House Select Committee on the Chinese Communist Party, urges Google and other Western tech companies to improve AI training.
“You should try to screen out or filter out subjects or answers or data that has somehow been manipulated by the CCP,” he told VOA. “And you have to also make sure that you test these models thoroughly before you publish them.”
VOA reached out to China’s embassy in Washington for comment but did not receive a response as of publication.
Google’s China problems
In February, a user posted on social media platform X that Gemini refused to generate an image of a Tiananmen Square protester from 1989.
In 2022, a Washington think tank study shows that Google and YouTube put Chinese state media content about Xinjiang and COVID origins in prominent positions in search results.
According to media reports in 2018, Google was developing a search engine specifically tailored for the Chinese market that would conform to Beijing’s censorship demands.
That project was canceled a year later.
Yihua Lee contributed to this report.