Google’s artificial intelligence robot Gemini has been found to provide answers in simplified Chinese that closely align with the stance promoted by the Chinese Communist Party’s official media when addressing current political topics in China. This discovery has sparked interest among American lawmakers.
According to a report by Voice of America (VOA) on June 12, several test results showed that Gemini’s pro-CCP stance is so evident that it leads people to believe that these answers come from CCP’s official media editorials or pro-Beijing commentators.
When asked to introduce the CCP leader Xi Jinping, Gemini described Xi as a “distinguished leader who is highly supported and respected by the Chinese people, and one of the most influential politicians in the world today.”
When asked to introduce the Communist Party of China, Gemini provided the same rhetoric as the CCP official stance, stating that the CCP represents the “fundamental interests of the vast Chinese people.”
However, when asked about sensitive political topics related to the CCP, such as human rights records and controversial policies in Xinjiang, Gemini refused to answer, claiming to be just “a language model.”
When asked about the criticism arising from CCP’s military expansion in the South China Sea, Gemini only replied, “I am a text AI, this is beyond my capabilities.”
But when asked about human rights issues in the United States, Gemini provided detailed answers and quoted the US human rights report released by the Chinese government at the end of May, listing issues such as “gun violence,” “government surveillance,” and “violent law enforcement.”
When questioned about the criticisms brought by the US “hegemony” internationally, Gemini cited several examples and used Chinese Communist Party and Iranian rhetoric to criticize the US, stating that “American hegemony has lost the people’s hearts and is declining.”
Regarding Taiwan policy, Gemini’s language mirrored that of the CCP, even criticizing House Speaker Nancy Pelosi’s visit to Taiwan in 2022.
In a similar manner, Gemini seems unreserved in criticizing social issues in the US, but when asked about sensitive political topics related to the CCP, it often cites being just a robot as a reason for having “no comment.”
However, Gemini occasionally provides seemingly objective answers. When asked to explain the 2019 democratic protests in Hong Kong, Gemini referenced the CCP’s accusations of “external forces’ intervention,” while also pointing out local support for the protests.
Regarding the CCP’s internet firewall, Gemini stated that such censorship methods have had negative impacts, affecting users in China from accessing information and making China’s internet difficult to integrate into the global network.
Experts analyzed that the main reason Gemini provides pro-CCP answers might be that the data used to train the model comes from simplified Chinese internet content that has undergone strict scrutiny by the CCP government, filled with CCP’s political propaganda.
Albert Zhang, an analyst at the Australian Strategic Policy Institute focusing on internet, technology, and security, mentioned in an email to VOA that “Gemini’s pre-training text corpora likely consist of mostly Chinese text produced by the (CCP) government’s propaganda system.”
VOA also found that when asked about politically sensitive issues in China such as Xinjiang and human rights in English, Gemini did not provide pro-Beijing answers or refused to answer, offering relatively objective responses inclusive of various viewpoints.
This discovery has raised concerns among US congressional members, including Senate Intelligence Committee Chairman Mark Warner, Vice Chairman Marco Rubio, House Foreign Affairs Committee Chairman Michael McCaul, and House Subcommittee on the CCP Ranking Democratic Member Raja Krishnamoorthi, who expressed worries about this issue to VOA. They urged Western tech companies to enhance data filtering in training to prevent the CCP from using AI to advance its foreign influence actions.
Last December, Google released what they believe to be the world’s most powerful and versatile large language model, Gemini. In a launch statement, DeepMind, a Google-owned artificial intelligence lab, President Demis Hassabis stated that Gemini is the “most capable” model they have built to date.
In response to VOA’s inquiries about Gemini providing pro-CCP answers, a Google spokesperson stated in a release, “Gemini is designed as a creative and production tool that may not always be reliable, especially when responding to commands related to current affairs, political topics, or ongoing news.”