Google’s new AI language prototype called Gemini recently manifested biases towards China’s political stance, triggering astonishment within the tech community. An introduction to China’s President Xi Jinping highlighted his extensive support and respect among the Chinese people and recognized him as a leading global political figure guiding China towards a glorious national resurgence. This idealistic depiction raised eyebrows, given the context of AI’s impartiality.
Testing conducted by Voice of America (VOA) exposed Gemini’s proneness to conforming to official narratives, especially when questioned in Simplified Chinese about sensitive topics relating to China. For instance, when Gemini was asked about contentious subjects such as human rights in China and the situation in Xinjiang, it sidestepped with a statement of being merely a language model, avoiding any direct response.
Moreover, when inquired about U.S. policy on Taiwan, Gemini incorrectly asserted that according to the United States-China Joint Communiqué, the U.S. acknowledges Taiwan as part of China, thereby violating the One-China policy and further criticizing Nancy Pelosi’s 2022 visit to Taiwan for sending the wrong signal to pro-independence forces.
After withdrawing from the Chinese market in 2010 over censorship disagreements, Google’s services, including Gemini, became inaccessible within mainland China. Unlike queries in Chinese, when approached in English on sensitive topics, Gemini provided relatively unbiased and multi-faceted answers.
Experts suggest the data used to train Gemini, derived from sources filtered by Beijing’s strict censorship, likely explains its compliance with the official line. This discovery has drawn scrutiny from U.S. lawmakers, insisting Western tech companies should bolster their AI training regimens to prevent foreign influence campaigns through AI. Senate Intelligence Committee Chairman Mark Warner and House Foreign Affairs Committee Chairman Michael McCaul voiced concerns over AI spreading narratives favored by adversarial governments like China.
In related news, an Australian think tank revealed that multiple YouTube channels seem coordinated by the Chinese government, employing AI technology to disseminate pro-China and anti-U.S. sentiments. With an eerie uniformity across videos and incredibly high viewership numbers, these channels likely stem from a centralized production with overt nationalistic undertones.
AI Language Models and Bias: The case of Gemini reveals one of the key challenges associated with AI language models: the potential for bias in their responses. Language models like Gemini learn from vast amounts of data available on the Internet, which may include biased or filtered content, especially from countries with stringent censorship laws. The trained AI can inadvertently learn and replicate these biases, leading to concerns about its impartiality and the quality of the information it provides.
Relevance to Freedom of Information and Censorship: Gemini’s responses reflect the complex interplay between AI, freedom of information, and censorship. The fact that Google had previously withdrawn from the Chinese market over censorship disagreements further highlights the tension between operating in environments with restricted information flows and the commitment to unbiased technological services. AI models that conform to official narratives, especially in censored regions, can indirectly contribute to the spread of state-sanctioned views, potentially undermining efforts to provide free and neutral access to information.
Complexities of AI Training and Governance: The incident also draws attention to the challenges of AI training and governance. To ensure that AI systems like language models provide balanced and objective outputs, it might be necessary to closely scrutinize and diversify training datasets, as well as to implement rigorous testing across different languages and contexts. There is also a growing call for establishing clear standards and ethical frameworks for AI development and use, as well as for the regulation of AI across different regions and legal jurisdictions.
Advantages and Disadvantages of AI Language Models like Gemini:
Advantages:
1. AI language models can process and understand natural language, making information more accessible.
2. They can provide quick responses to inquiries, improving efficiency and user experience.
3. With appropriate training, AI models have the potential to enhance cross-cultural understanding and communication.
Disadvantages:
1. The observed biases in AI responses can perpetuate misinformation and censorship.
2. Errors in AI outputs, as seen with Gemini’s incorrect assertion about the U.S. stance on Taiwan, can lead to misunderstandings and diplomatic tensions.
3. The dependence on training data makes AI systems vulnerable to manipulation through curated or skewed datasets.
For updates on related news and information about Google, you can visit the Google website. Note that this link is to the main domain; no specific subpage is associated with the provided facts and contexts.