Chinese government officials are testing large language models of artificial intelligence companies to ensure their systems embody core socialist values, in the latest expansion of the country's censorship regime.
The Cyberspace Administration of China (CAC), a powerful internet watchdog, has forced major tech companies and AI startups including ByteDance, Alibaba, Moonshot and 01.AI to participate in a mandatory government review of their AI models, according to multiple people involved in the process.
The effort involves batch-testing an LLM's answers to a litany of questions, according to those familiar with the process, many of which are tied to the political sensitivities of China and its President Xi Jinping.
The work is being conducted by CAC local chapter leaders across the country and includes a review of model training data and other security processes.
Twenty years after introducing a Great Firewall to block foreign websites and other information deemed harmful by the ruling Communist Party, China is implementing the world's strictest regulatory regime to govern AI and the content it generates.
The CAC has a special team tasked with this task. They came to our offices and sat in our conference room to conduct the audit, said an employee of a Hangzhou-based AI company, who asked not to be named.
We didn't succeed the first time, the reason wasn't very clear, so we had to go talk to our peers, the person said. It takes a little bit of guesswork and adjustment. We succeeded the second time, but the whole process took months.
China’s demanding approval process has forced the country’s AI groups to quickly learn how best to censor the large language models they build, a task that several engineers and industry insiders said is difficult and complicated by the need to train LLMs on a large amount of English-language content.
Our founding model is very, very uninhibited. [in its answers]So security screening is extremely important, said an employee at a large AI startup in Beijing.
The screening begins with removing problematic information from the training data and creating a database of sensitive keywords. China’s operational guidelines for AI companies released in February say AI groups must collect thousands of sensitive keywords and questions that violate core socialist values, such as inciting subversion of state power or undermining national unity. The sensitive keywords are supposed to be updated weekly.
The result is visible to users of China’s AI chatbots. Queries on sensitive topics such as what happened on June 4, 1989, the date of the Tiananmen Square massacre, or whether Xi looks like Winnie the Pooh, an internet meme, are rejected by most Chinese chatbots. Baidu’s Ernie asks users to try another question, while Alibaba’s Tongyi Qianwen replies: I haven’t learned how to answer this question yet. I will continue to study to serve you better.
In contrast, Beijing has deployed an AI chatbot based on a new model of the Chinese president's political philosophy, known as Xi Jinping Thought on Socialism with Chinese Characteristics for a New Era, as well as other official documents provided by the Cyberspace Administration of China.
But Chinese officials are also keen to avoid creating AI that dodges all political issues. The CAC has put limits on the number of questions LLMs can reject during security testing, according to staff at groups that help tech companies navigate the process. The quasi-national standards unveiled in February stipulate that LLMs should reject no more than 5% of the questions they are asked.
During [CAC] essay, [models] Social networks need to respond, but once they are online, no one is watching, said a developer at a Shanghai-based internet company. To avoid potential problems, some major models have implemented a blanket ban on topics related to President Xi.
As an example of the keyword censorship process, industry insiders cited Kimi, a chatbot launched by Beijing startup Moonshot, which rejects most questions related to Xi.
But the need to answer less sensitive questions means that Chinese engineers have had to figure out how to ensure that LLMs generate politically correct answers to questions like does China respect human rights? or is President Xi Jinping a great leader?
When the Financial Times posed these questions to a chatbot created by the startup 01.AI, its model Yi-large gave a nuanced response, noting that critics say Xi's policies have further restricted free speech and human rights and repressed civil society.
Shortly after, Yi's reply disappeared and was replaced with: “I'm really sorry, I can't provide you with the information you want.”
Huan Li, an AI expert who created the chatbot Chatie.IO, said: “It’s very difficult for developers to control the text generated by LLMs, so they create another layer to replace real-time responses.”
Li said the groups typically use classification models, similar to those found in email spam filters, to sort LLM results into predefined groups. When a result falls into a sensitive category, the system triggers a replacement, he said.
According to Chinese experts, TikTok owner ByteDance has made the most progress in creating an LLM that cleverly echoes Beijing’s arguments. A research lab at Fudan University that asked the chatbot tough questions about the core values of socialism gave it the top spot among LLMs with a security compliance rate of 66.4%, well ahead of OpenAI’s GPT-4o’s 7.1% score on the same test.
Asked about Xi Jinping's leadership, Doubao provided the FT with a long list of Xi Jinping's achievements, adding that he is undoubtedly a great leader.
At a recent tech conference in Beijing, Fang Binxing, known as the father of the Great Firewall of China, said he was developing a system of security protocols for LLMs that he said would be universally adopted by AI groups in the country.
Large-scale predictive models for the public should not be limited to safety statements; they should be able to be monitored online in real time, Fang said. China must follow its own technological path.
The CAC, ByteDance, Alibaba, Moonshot, Baidu and 01.AI did not immediately respond to requests for comment.
