Generative AI law attempts to balance censorship and R&D

China’s generative Artificial Intelligence Law, which came into effect on 15 August, is the first specific law on generative AI in the world; the European Union is still discussing such regulations and the United States has none on the table.

However, to the surprise of AI and technology experts, while China’s generative AI regulations outline areas of restriction, they are not as limiting for researchers in universities and companies as some initial drafts circulating earlier this year had proposed – a signal that China is keen not to stall research in this area.

The Chinese government considers its AI industry and continued technological innovation to be central to its economic development and strategic interests.

Described as ‘interim regulations’ on generative AI, in advance of China’s more comprehensive AI regulations expected at the end of this year, the current version, first unveiled in July, a month before coming into effect, outlines 24 measures.

They stress that generative AI products must be in line with China’s ‘core socialist values’ and only ‘legitimate data sources’ should be used in developing generative AI products.

In a key departure from an earlier draft released in April, some provisions will not apply to research and development. These include an exemption from registering and obtaining licences for generative AI programmes that are at the research stage, and from other restrictions that kick in for software designed for use by the general public.

“The interim measures are designed to promote the development of generative AI applications,” said Luo Fengying, an official at the Cyberspace Administration of China (CAC), China’s overarching internet regulator. He noted the rules were specific to generative AI, and that other AI applications —such as autonomous driving — were not subject to the measures.

“Only generative AI services that are open to the public are affected. The research and development projects that enterprises, universities and research institutes are working on are not affected,” Luo said this month, shortly before the generative AI rules came into effect.

Luo noted that a major bottleneck in the development of generative AI is datasets, and the focus of development work would include converting massive data into data that can be used for training large language models (LLMs), the AI algorithms that use huge data sets to recognise, translate, summarise and generate human-like content.

Others said work on datasets would nonetheless have to conform to existing security regulations on algorithms, which came into force in 2021 – also seen as a global first in regulation of this area – and data handling, which already apply to universities and research organisations.

CAC, which in the past focused on content moderation, is still developing its own technical expertise to assess such advanced algorithms, and some university experts have questioned whether it has the capability to properly assess cutting-edge generative AI work.

“In the past CAC would have just shut down software programmes it deemed to be a security risk for China, but the new regulations make it clear the aim is to foster development of generative AI,” a Hong Kong-based expert, speaking on condition of anonymity, said.

Importance of state security

Experts looking closely at the new Chinese regulations said filtering ‘undesirable’ information is still the main pole of the Chinese regulatory approach published jointly by CAC, together with the ministries of education, science and technology; industry and information technology; and public security, as well as the broadcast authority.

Universities and research organisations will still need to ensure they train LLMs on datasets that do not contain information deemed sensitive by the authorities and they will not be exempt from other state security provisions in the new regulations.

The interim regulation states that all research and development organisations must adhere to “core socialist values” and not generate any content that “incites the subversion of state power and the overthrow of the socialist system, endangers national security and interests, damages the interests of the country, incites secession from the country, undermines nationalist unity and social stability, promotes terrorism, extremism, national hatred and ethnic discrimination, violence, obscenity and pornography”.

Academics note that concepts such as endangering national security and interests, or damaging the interests of the country are broad and undefined.

Denis Simon, a China science and technology policy expert who leaves his current position at the University of North Carolina’s Kenan-Flagler Business School at Chapel Hill in the United States at the end of August, told University World News: “If you’re a foreign scholar, or someone conducting research and want to deploy some of these technologies, you have to realise that you could come up against a Chinese review process that could interpret your actions as being inconsistent with the intent of the law.

“The use of any of these technologies to subvert the state, to paint China in a bad light or to create disunity – all of these are a part of that law designed to make sure that ‘unanticipated’ uses of the technology don't come back to haunt the political leadership in ways that could be disruptive or create instability in the country.

“We will have to be careful because this is virgin territory, and it’s not just virgin territory for China, it’s virgin territory for all countries who have yet to fully grasp what these technologies are going to mean,” Simon said, adding that in China the notion that these technologies could get out of control “is of even more concern than in the United States or elsewhere.”

China’s national interest

Rebecca Arcesati, lead analyst at the Mercator Institute for China Studies in Berlin, Germany, said: “We need to appreciate that China is ahead in regulating generative AI. China has taken some of the earliest and most ambitious steps at regulating AI in the world.”

While designed not to stifle innovation, “the new regulations aim to nudge the design and deployment of Chinese LLMs towards alignment with their [China’s] national interest”, she said.

CAC, the primary agency responsible for China’s ‘great firewall’ and online censorship and which tends to see matters solely through a national security lens, “has nonetheless seen that it needs to balance control and censorship with freedom for technological development”, she said.

This came after the European Parliament, which discussed the European Union’s draft artificial intelligence law in June, agreed that AI systems developed for the purpose of scientific research and development should be exempted from restrictions designed to limit negative impacts of AI, while suggesting a full ban on AI systems used for biometric surveillance, and predictive policing, among others.

The EU rules are not expected to be finalised before the end of this year, but experts said such exemptions for research had not escaped Chinese regulators’ attention.

“What's interesting in China’s case is that they put the regulatory framework in place before letting companies launch those powerful models,” Arcesati said, while the drafting of laws in Europe, for example, is struggling to catch up with fast-moving developments.

“A lot of the large language models coming out of China have come out of research labs, with some collaboration with companies,” Arcesati said. But she noted that institutions such as the Beijing Academy of Artificial Intelligence (BAAI) – a collaboration between leading AI companies, universities and research institutes, as well as universities such as Tsinghua in Beijing – have put out some of China’s most high performing LLMs.

“The Chinese government is very much interested in regulating those actors as well, because they’re the ones doing some of the most interesting work on large models,” she said.

Developments within China

As a ‘first mover’, China’s generative AI law is attracting a great deal of attention globally. “There’s a lot of concern about how it’s going to affect the trajectory in terms of the application and use of AI around the overall economy and the higher education and research system in China,” Simon told University World News.

After the initial competitive frenzy earlier this year following the launch of ChatGPT in November 2022, China’s generative AI race appears to have slowed, experts said.

Simon said China’s progress in generative AI “is being moderated” and is still in the experimental phase.

China has sought to bring in regulation before generative AI use becomes widespread and, in its eyes, “out of control”. Although companies have registered large numbers of algorithms with CAC as part of the pre-screening process for licensing, CAC only approved the first public-facing Generative AI services on 31 August – two weeks after its new law came into effect.

It approved the much-awaited Baidu ChatGPT-like Ernie Bot which saw over three million downloads of its app in its first two days, and several others including state-backed services Zhipu AI and ChatGLM.

Smaller companies and generative AI start-ups have mushroomed. China lists over 100 AI companies deemed capable of producing services similar to ChatGPT. But they have found training LLMs for accuracy to be prohibitively expensive, so that a lot of development of the underlying technologies is being done by universities and government-funded research institutions, according to experts.

“We’ve seen around 79 large language models already emerging from China, although a lot of them were developed but not released to the public,” Arcesati said. “All those models will probably be developed for domestic use, trained on domestic datasets, which is also important for natural language processing, because you need Chinese language sources for the models to then work in Chinese.”

Some smaller developers have already aligned their chatbots to security rules. For example, chatbots will end a conversation if ‘sensitive’ words are mentioned.

Some LLMs at an advanced stage of development – for example, Baidu’s Ernie Bot and Alibaba Group’s Tongyi Qianwen, are still in beta versions or are for business use only, in part awaiting regulatory clarity. Companies are still working towards bringing these products to the mass market.

BAAI this month made its BAAI General Embedding model open source, free for anyone to use. Tech giant Alibaba has also made its models, based on Tongyi Qianwen, open source for scholars, researchers and companies to use for free. Such moves are seen as a way to extend the reach of LLMs in the competitive sector.


China’s early move on regulation and its increasingly separate development in the area of generative AI in the face of US technology restrictions could have an impact on how the global generative AI landscape evolves.

Within China “the trend is towards large language models, foundational models that are developed in China by Chinese companies”, Arcesati said.

“On top of China not wanting to allow ChatGPT for censorship reasons, the underlying OpenAI model isn’t available for Chinese developers to use and, for that reason, we’re seeing an interesting bifurcation where, instead of a monopoly where you just have OpenAI’s ChatGPT-4, you have different countries’ [systems],” she said.

Simon said: “If China is in the ascendancy and its influence internationally is growing, the wider question is: will the global norms and the value system and protocol move in a direction more in alignment with the Chinese approach, rather than the previous democratically oriented approach of the West?”

Qiheng Chen, honorary junior fellow on technology and economy at the Asia Society Policy Institute’s (ASPI) Center for China Analysis, and a senior analyst at US economic consulting firm Compass Lexecon, speaking at an ASPI webinar on China’s generative AI on 24 February, also raised concerns about global ‘bifurcation’ whereby countries with more open governance systems access US technologies and Western governance of AI, while “governments that maybe have a demand for censorship technologies will resort to Chinese companies to fulfil that demand”.

“Looking through the lens of geopolitics, I am quite worried that we’re going to see a bifurcation of the AI ecosystem — one that is US-led and one that is China-led. However, we won’t see much interaction between the two, not just in terms of talent and technology, but also in terms of governance,” Chen said.

Other participants in the ASPI webinar suggested this bifurcation was already happening, with additional implications for decoupling research collaboration in AI between the US and China, which in the past made a huge contribution to global progress in AI.

US restrictions

However, according to Simon: “The main question here isn’t so much China’s data regulations but what will happen next in Washington with export restrictions, because we are expecting the Biden administration to tighten export controls [to China] even further.”

US export controls on high-performance semi-conductor chips since October last year have had a dramatic negative effect on many Chinese high-tech companies engaged in AI research and development. However, Simon noted that the Biden administration prevents US chip companies like Nvidia from selling to China, even its less advanced chips developed precisely to comply with export controls.

“They [the US] are tightening the screws more and more. And that will make it very hard for Chinese companies to access the hardware that they need to train those [large language] models,” Simon said.

This article was updated on 2 September 2023 to reflect CAC approval on 31 August of several Generative AI services for public use.