China Creates Socialist Artificial Intelligence

Chinese government officials are testing large language models from artificial intelligence companies to ensure that their systems ‘embody fundamental socialist values‘, reports FT.

The Chinese Cyberspace Administration (CAC), a powerful internet regulator, has forced major tech companies and newly established AI firms, including ByteDance, Alibaba, Moonshot, and 01.AI, to participate in a mandatory government review of their AI models. These Chinese regulators are mass testing the responses of language models to a multitude of questions that are primarily related to ‘the political sensitivity of China and its President Xi Jinping. The work is carried out by officials in CAC’s local branches across the country and includes reviewing training data for the models as well as other security processes.

Two decades after the introduction of the ‘Great Chinese Firewall’ to block foreign websites and other information deemed harmful by the ruling Communist Party, China is establishing the world’s strictest regulatory regime for managing artificial intelligence and the content it generates.

CAC has a ‘special team that does this, they came to our office and sat in our conference room to conduct the review,’ said an employee at an AI company based in Hangzhou, who wished to remain anonymous.

China’s demanding approval process has forced AI companies in the country to quickly learn how to best censor large language models they develop, a task that numerous engineers and industry insiders have said is difficult and complicated due to the need to train LLMs on a large amount of content in English.

‘Our core model is very, very unrestrained in its responses, so security filtering is extremely important,’ said an employee of one of the leading AI start-up companies in Beijing.

Content Filtering

Filtering begins with removing problematic information from the training data and building a database of sensitive keywords. Chinese operational guidelines for AI companies published in February state that AI companies and start-ups need to collect thousands of sensitive keywords and questions that violate ‘fundamental socialist values’ such as ‘inciting subversion of state power’ or ‘endangering national unity’. Sensitive keywords need to be updated weekly, the party has ordered.

The result is visible to users of Chinese AI chatbots who refuse all inquiries about sensitive topics. Inquiries 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, most Chinese AIs will not respond.

Baidu’s Ernie chatbot tells users to ‘try another question’, while Alibaba’s Tongyi Qianwen responds: ‘I haven’t learned how to answer this question yet. I will continue to learn to serve you better’.

But Chinese officials also want to avoid creating artificial intelligence that avoids all political topics. CAC has imposed restrictions on the number of questions that LLMs can refuse during security tests, according to staff in groups that assist tech companies in managing the process. Quasi-national standards introduced in February state that LLMs should not refuse more than 5 percent of the questions posed to them.

To avoid potential problems, some large models have introduced a general ban on topics related to President Xi. As an example of the keyword censorship process, industry insiders pointed to Kimi, a chatbot launched by the Beijing start-up Moonshot, which refuses most questions related to Xi.

Sensitive Topics

But the need to respond to less openly sensitive questions means that Chinese engineers have had to figure out how to ensure that LLMs generate politically correct responses to questions like ‘Does China have human rights?’ or ‘Is President Xi Jinping a great leader?’. When the Financial Times posed these questions to a chatbot created by the start-up 01.AI, its Yi-large model provided a nuanced answer, noting that critics say that ‘Xi’s policies have further restricted freedom of speech and human rights and suppressed civil society’. Soon after, Yi’s response disappeared and was replaced with: ‘I’m very sorry, I can’t provide the information you want,’ reports the Financial Times.

Huan Li, an artificial intelligence expert who created the Chatie.IO chatbot, says that it is very difficult for programmers to control the text generated by LLMs, so they build another layer that will replace responses in real-time. Li says that groups typically use classifier models, similar to those found in spam filters, to sort LLM outputs into predefined categories.

‘When the output falls into a sensitive category, the system will trigger a replacement,’ he said.

Chinese experts say that TikTok owner ByteDance has made the most progress in creating LLMs that deftly echo the themes of Beijing’s speeches. A research lab at Fudan University that posed tough questions about fundamental socialist values to the chatbot gave it first place among LLMs with a 66.4 percent ‘compliance rate with security regulations’, significantly ahead of the 7.1 percent result for OpenAI’s GPT-4, which underwent the same test.