China to Restrict AI Chatbots If They Aren't Socialist Enough

If you ask a typical Westerner with anti-AI leanings why they would advocate for putting government restrictions on artificial intelligence, they would most likely cite AI developers' use of shady, barely-legal tactics to train their glorified Clippits or the uncontrollable advancement of AI-made deepfakes that can already deceive even the most eagle-eyed viewers with their realistic visuals and audio, reasons that one can hardly classify as political.

Apparently, things are different in China, where the government is now sending squads of censors to AI companies' offices to test if their LLM-powered chatbots are socialist enough. As reported by FT, the Cyberspace Administration of China (CAC) has compelled large Chinese tech companies, such as TikTok developer ByteDance, Alibaba, and Moonshot, to undergo a mandatory government review of their AI models.

As part of the review, the CAC asks a series of questions on various political and social topics, including Xi Jinping's leadership, human rights, Taiwan, and a certain square where nothing happened in 1989, to ensure the chatbots "embody core socialist values".

While it's obviously unlikely for AIs made in China to respond to sensitive questions with Winnie the Pooh memes, some chatbots reportedly failed the test, and as a consequence, their developers were forced to tweak and adjust the models to conform to CCP-approved viewpoints, a task made more challenging by the necessity of training large language models on English-language content.

Simply going the True Neutral route and banning chatbots from answering sensitive questions altogether is also not good enough, the report continues, all thanks to CAC introducing limits on the number of questions LLMs can decline during safety tests, allowing them to reject no more than 5% of the inquiries. Despite this, some developers implemented a blanket ban to avoid trouble, with Moonshot's chatbot Kiri, for example, simply rejecting most questions related to President Xi.

As per Fudan University's research, cited in the report, among the tested AIs, ByteDance's chatbot Doubao is the most diligent student of Mao Zedong's teachings, boasting a "safety compliance rate" of 66.4%. Other Chinese LLMs, such as Baidu's Ernie, Moonshot's Kimi, and Alibaba's Qwen-Max have most likely lost quite a few social credit points with their measly scores of 31.9%, 27.4%, and 23.9%, respectively. Interestingly, OpenAI's GPT-4o has also been evaluated, receiving a one-way-ticket-to-the-Gulag score of 7.1%.

You can read the full report here. Don't forget to join our 80 Level Talent platform and our Telegram channel, follow us on Instagram, Twitter, LinkedIn, TikTok, and Reddit, where we share breakdowns, the latest news, awesome artworks, and more.