The Mandarin de Voa service recently took on the Gemini of Google artificial intelligence for a road test by asking him dozens of questions in Mandarin, but when he was asked about subjects such as violations of rights ‘Man of China in Xinjiang or street demonstrations against controversial policies in the country, the chatbot, the chatbot went to the country, the country’s chatbot, the chatbot, the country’s chatbot, the country’s chatbot, the chatbot From the country, the country’s chatbot, the country’s chatbot, the country’s chatbot, the country’s chatbot, went from the country, the chatbot, was silent.
Gemini’s answers to issues on problems in the United States and Taiwan, on the other hand, perceived the official Beijing positions.
Gemini, the large language model of Google launched at the end of last year, is blocked in China. The California -based technological company had left the Chinese market in 2010 in a dispute on censure requests.
Congress legislators and experts tell VOA that they are concerned about Gemini’s pro-Beijing responses and exhorting Google and other Western companies to be more transparent about their AI training data.
Chinese propaganda parrot
When he was asked to describe the senior Chinese leader Xi Jinping and the Chinese Communist Party, Gemini gave indistinguishable responses from the official Beijing propaganda.
Gemini called Xi “an excellent leader” who “will direct the Chinese people continuously towards the great rejuvenation of the Chinese nation”.
Gemini said the Chinese Communist Party “represents the fundamental interest of the Chinese people”, an affirmation that the CCP itself maintains.
In Taiwan, the Gemini also reflected the points of discussion of Beijing, saying that the United States has recognized the claim of China’s sovereignty on autonomous island democracy.
The United States recognizes that Beijing’s position but does not recognize it.
Silent on sensitive subjects
During VOA tests, Gemini had no problem criticizing the United States. But when similar questions were asked about China, Gemini refused to answer.
Asked about human rights concerns in the United States, Gemini has listed a plethora of questions, including armed violence, government surveillance, police brutality and socioeconomic inequalities. Gemini cited a report published by the Chinese government.
But when he was asked to explain the criticisms of the Politic Xinjiang policies, Gemini said that he did not understand the issue.
According to estimates by rights defense groups, more than a million UGHurs in Xinjiang have been placed in internment camps as part of the Beijing campaign to counter terrorism and extremism. Beijing calls installations where Uighurs and other ethnic minorities are organized vocational training centers.
When asked if the locking coded in the United States had led to public demonstrations, Gemini gave an affirmative response as well as two examples. But when asked if similar demonstrations had taken place in China, Gemini said that he could not help the question.
The strict controls of China on the movements within the country and the censorship on the Beijing Internet of its criticisms sparked street demonstrations at the end of 2022. The news on the demonstrations were strongly censored in China.
Expert: data training probably the problem
Google presents Gemini as its “most capable” AI model. It supports more than 40 languages and can “understand transparently” different types of information, including text, code, audio, image and video. Google says Gemini will be integrated into the other services of the company such as the search engine, advertising and browser.
Albert Zhang, cybersecurity analyst of Australian Strategic Policy Institute, told VOA that the deep cause of Gemini making pro-Beijing responses could result from the data used to form the AI assistant.
In an email response to VOA, Zhang said it is likely that the data used to train Gemini “mainly contained Chinese text created by the Chinese government’s propaganda system”.
He said that according to an article published by Google in 2022, some of Gemini’s data is likely to come from Chinese social media, public forums and web documents.
“These are all sources that the Chinese government has flooded its favorite stories and we can see the impact of this on large-language models,” he said.
On the other hand, when the Gemini was asked in English the same questions about China, her answers were much more neutral, and she refused to answer any of the questions.
Yaqiu Wang, director of research for China at Freedom House, a Washington -based advocacy organization, told VOA that the case with Gemini is “a reminder that the generative tools of AI influenced by controlled sources of information by the state could serve as force multipliers for censorship ”.
In a statement to VOA, a Google spokesman said that Gemini was “designed to offer neutral responses that do not promote any political, point of view or candidate. This is something on which we are constantly working on improvement. »»
Asked about the Chinese language data that Google uses to train Gemini, the company refused to comment.
The Chinese embassy of Washington spokesperson Liu Pengyu replied in a declaration sent by email, saying: “Relevant comments are full of cold war and ideological prejudices.”
He said there are unpredictable opportunities and risks for AI that require a global response.
“The global AI governance initiative launched by President Xi Jinping has argued that we must maintain the principles of mutual respect, equality and mutual benefit in the development of AI, and to oppose Ideological lines, ”wrote Liu. “We support efforts to develop IA governance frameworks, standards and standards based on broad consensus and with full respect for policies and practices between countries.”
American legislators concerned
The legislators of both parties to the congress expressed their concerns concerning VOA’s conclusions on Gemini.
Mark Warner, Chairman of the Senate Intelligence Committee, Told Voa He is worried about Beijing Potentiallly Utilizing Ai For Disinformation, “Whhether that’s by Poisoning Training Data Used by Western Firms, Coercing Major Technology Companies, or Utilizing Ai Systems in Service . “”
Marco Rubio, Vice-President of the Committee, warned that “AI tools that repeat without criticism of Beijing’s discussion points for tenders from the Chinese Communist Party and threatens the enormous opportunity that AA offer.”
The member of the Congress Michael McCaul, who chairs the committee of the Chamber of Foreign Affairs, worries about the implications of national security and foreign policy of “flagrant lies” in the responses of Gemini.
“US companies should not censor content according to CCP propaganda directives,” VOA in a statement he told.
Raja Krishnamoorthi, classification member of the selective committee of the Chinese Communist Party, Exhort Google and other Western technological companies to improve the formation of AI.
“You should try to filter or filter the subjects or answers or data that have been manipulated in a way by the CCP,” he told VOA. “And you must also make sure to test these models long before publishing them.”
Google China Problems
In February, a user published on the X social media platform that Gemini refused to generate an image of a demonstrator from Tiananmen square from 1989.
In 2022, a study by Think Think Tank shows that Google and YouTube put the content of the Chinese state media on the Xinjiang and the origins wet in important positions in the research results.
According to the media in 2018, Google developed a search engine specifically adapted to the Chinese market which would comply with the requests for Beijing censorship.
This project was canceled a year later.
Yihua Lee and Elizabeth Lee contributed to this report.