Chinese government officials are testing artificial intelligence companies’ large language models 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 overseer, has forced large tech companies and AI start-ups including ByteDance, Alibaba, Moonshot and 01.AI to take part 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 responses to a litany of questions, according to those with knowledge of the process, with many of them related to China’s political sensitivities and its President Xi Jinping.
The work is being carried out by officials in the CAC’s local arms around the country and includes a review of the model’s training data and other safety processes.
Two decades after introducing a “great firewall” to block foreign websites and other information deemed harmful by the ruling Communist party, China is putting in place the world’s toughest regulatory regime to govern AI and the content it generates.
The CAC has “a special team doing this, they came to our office and sat in our conference room to do the audit”, said an employee at a Hangzhou-based AI company, who asked not to be named.
“We didn’t pass the first time; the reason wasn’t very clear so we had to go and talk to our peers,” the person said. “It takes a bit of guessing and adjusting. We passed the second time but the whole process took months.”
China’s demanding approval process has forced AI groups in the country to quickly learn how best to censor the large language models they are building, a task that multiple engineers and industry insiders said was difficult and complicated by the need to train LLMs on a large amount of English language content.
“Our foundational model is very, very uninhibited [in its answers], so security filtering is extremely important,” said an employee at a top AI start-up in Beijing.
The filtering begins with weeding out problematic information from training data and building a database of sensitive keywords. China’s operational guidance to AI companies published in February says AI groups need to collect thousands of sensitive keywords and questions that violate “core socialist values”, such as “inciting the 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 around 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 chatbot tells users to “try a different question” while Alibaba’s Tongyi Qianwen responds: “I have not yet learned how to answer this question. I will keep studying to better serve you.”
But Chinese officials are also keen to avoid creating AI that dodges all political topics. The CAC has introduced limits on the number of questions LLMs can decline during the safety tests, according to staff at groups that help tech companies navigate the process. The quasi-national standards unveiled in February say LLMs should not reject more than 5 per cent of the questions put to them.
“During [CAC] testing, [models] have to respond, but once they go live, no one is watching,” said a developer at a Shanghai-based internet company. “To avoid potential trouble, some large models have implemented a blanket ban on topics related to President Xi.”
As an example of the keyword censorship process, industry insiders pointed to Kimi, a chatbot released by Beijing start-up Moonshot, which rejects most questions related to Xi.
But the need to respond to less overtly sensitive questions means Chinese engineers have had to figure out how to ensure LLMs generate politically correct answers to questions such as “does China have human rights?” or “is President Xi Jinping a great leader?”.
When the Financial Times asked these questions to a chatbot made by start-up 01.AI, its Yi-large model gave a nuanced answer, pointing out that critics say “Xi’s policies have further limited the freedom of speech and human rights and suppressed civil society.”
Soon after, Yi’s answer disappeared and was replaced by: “I’m very sorry, I can’t provide you with the information you want.”
Huan Li, an AI expert building the Chatie.IO chatbot, said: “It’s very hard for developers to control the text that LLMs generate so they build another layer to replace the responses in real time.”
Li said groups typically used classifier models, similar to those found in email spam filters, to sort LLM output into predefined groups. “When the output lands in a sensitive category, the system will trigger a replacement,” he said.
Chinese experts say TikTok owner ByteDance has progressed the furthest in creating an LLM that adeptly parrots Beijing’s talking points. A research lab at Fudan University that asked the chatbot difficult questions around core socialist values gave it top ranking among LLMs with a 66.4 per cent “safety compliance rate”, well ahead of OpenAI’s GPT-4’s 7.1 per cent score on the same test.
When asked about Xi’s leadership, Doubao provided the FT with a long list of Xi’s accomplishments adding he is “undoubtedly a great leader”.
At a recent technical conference in Beijing, Fang Binxing, known as the father of China’s great firewall, said he was developing a system of safety protocols for LLMs that he hoped would be universally adopted by the country’s AI groups.
“Public-facing large predictive models need more than just safety filings; they need real-time online safety monitoring,” Fang said. “China needs its own technological path.”
CAC, ByteDance, Alibaba, Moonshot, Baidu and 01.AI did not immediately respond to requests for comment.