China Lays Out Strict Rules for ChatGPT-Like AI Tools

The Cyberspace Administration of China will require companies to go through a government security review before providing such services and make companies responsible for the content their AI services generate, according to a draft of the rules. Content generated by such services shouldn’t contain elements that could subvert state power, incite secession or disrupt social order, the rules state.

The proposals show China’s leaders want to maintain a strong hand in regulating new technology even as it eases up on a two-year crackdown on the country’s internet companies.

Governments around the world are discussing whether and how to govern the new wave of generative AI tools, with some of the world’s leading tech figures warning they could be used to spread harmful information or discriminate. In the U.S., the Biden administration has begun examining whether checks need to be placed on the tools. Italy temporarily banned ChatGPT, saying the AI chatbot has improperly collected and stored information.

The Chinese regulator’s announcement came the same day that

Alibaba

BABA -1.17%

Group Holding Ltd. rolled out its large language model, called Tongyi Qianwei, which it plans to integrate across products including its search engine and voice assistant, as well as entertainment and e-commerce.

A day earlier,

SenseTime Group Inc.,

best known for surveillance products such as facial-recognition systems, launched a ChatGPT-like service, SenseChat, and a cluster of apps based on its large AI model system SenseNova. Huawei Technologies Co. on Saturday said it has rolled out services based on Pangu, a collection of large AI models that it has been developing since 2019, to enterprise clients in industries including finance, pharmaceuticals and meteorology.

While Chinese tech companies have said their AI products still lag behind OpenAI’s ChatGPT, which is unavailable in the country, they are pushing to develop their own versions of the technology. Generative AI could open new revenue opportunities for them after the internet sector was battered in recent years by tighter regulatory oversight and a sluggish economy.

The companies have to navigate through constraints including U.S. curbs on them buying advanced chips that are required for training AI models, as well as China’s rigid censorship policies.

China’s proposed rules are more detailed than general guidelines that are being discussed in other jurisdictions, said You Chuanman of the Chinese University of Hong Kong, Shenzhen, who specializes in tech regulation and global governance. China’s rules would include a prohibition on profiling users and impose strict controls on AI-generated content, which could jeopardize innovation as China battles the U.S. in developing advanced technologies, he said.

“AI is a challenge for global governance,” Dr. You said. “Governments from different countries should work together to deliver a global standard.”

According to the draft rules, companies would be responsible for protecting users’ personal information, while data that developers use to train their AI products should be compliant with Chinese laws.

China’s censorship rules not only restrict the content AI could produce but limit the materials Chinese developers can use to train their products to a much smaller pool than their foreign rivals, industry observers say.

China’s internet regulator didn’t state when the rules, which are open for public consultation until May 10, would be rolled out.

Last month, China’s search-engine company

Baidu Inc.

became the country’s first internet company to launch an AI-powered chatbot, Ernie Bot.

More Chinese tech companies have followed, seeking to commercialize AI technologies that have cost hundreds of millions of dollars to develop.

Alibaba has been building its own large language models since 2019 and last September launched its proprietary system Tongyi that groups together several generative AI models.

It will first integrate Tongyi Qianwen, which translates as “thousands of questions for universal truths,” into its workplace messaging app DingTalk. It could be used to summarize meeting notes as well as draft emails and business proposals. The AI model will also be embedded into Alibaba’s voice assistant Tmall Genie, including being able to provide recipes and offer travel tips, it said.

Alibaba’s chatbot is able to generate text, lines of computer code and do mathematics, but hasn’t allowed users to produce pictures yet. Alibaba said it is capable of handling tasks in both Chinese and English.

During a live demonstration Monday, SenseTime’s chatbot, designed to mainly handle Chinese-language questions, was able to produce text, computer code and images. The company also used the service to summarize academic papers and offer medical advice. Hong Kong-based SenseTime is backed by Alibaba and

SoftBank Group Corp.

“Having a model itself isn’t sufficient,” said Boris Van, an analyst at Bernstein Research, referring to Chinese tech companies’ AI efforts more generally. “You need to have products and applications to incorporate the model and use the model to deliver actual things that customers want.”

Write to Raffaele Huang at raffaele.huang@wsj.com