Unpacking China's Generative AI Draft Regulations
5 curated resources on the recent landmark regulations for AI-generated content
As you probably know, the Cyberspace Administration of China (CAC) published draft regulations for generative AI services last month.1
This is, of course, not China’s first set of regulations targeting AIGC (AI-generated content). The earlier deep synthesis regulations, known for specifically targeting “deepfake” technology, do cover the development and use of AI-generated text, speech, video, and more.2 It was these regulations that came into play in the recent arrest of a Gansu man who allegedly used ChatGPT to generate and disseminate false news (the sixth article of those regulations explicitly forbids this exact kind of thing).
In contrast to the deep synthesis regulations, which target a narrow scope of AIGC primarily consisting of deceptive content, the generative AI draft regulations more comprehensively target AIGC services, such as ChatGPT-like chatbots.
The deadline for feedback on these draft regulations was last Wednesday, May 10. There hasn’t been an official announcement regarding when these regulations will go into effect, but given the increasing mainstream use of AIGC, we’ll likely hear some concrete news soon.
Since the announcement of the draft regulations, there’s been a wealth of English- and Chinese-language analysis of these policies and what they could mean for the development of China’s AIGC industry. Rather than throw another piece of long-form analysis onto the pile, I’ve decided to publish a roundup of some of the most insightful pieces I’ve come across on the draft regulations.
(This list is far from exhaustive, so feel free to post links to additional pieces in the comments.)
A Quick Reading List on China & AI
Analyzing the AI Draft Regulations
5 Recommendations for Improving the Generative AI Draft Regulations (Chinese / English)
A virtual roundtable including members of Peking University Law School and CAICT, this piece (originally published by 数字经济与社会 on WeChat and translated into English here) discusses the draft regulations and proposes specific revisions.
Examples include clarifying the scope of the AIGC “providers” mentioned in the regulations and increasing legal requirements for the sources of data used to train AI models.
Credit goes to Matt Sheehan for spotting the original Chinese piece and to Jeffrey Ding for spearheading the translation in his newsletter earlier this month.
DigiChina Forum: How Will China’s Generative AI Regulations Shape the Future? (English)
This is an informative collection of takes on the generative AI regulations, featuring insightful opinions from Rogier Creemers, Paul Triolo, and over half a dozen others in the China policy field.
One highlight: Yan Luo and Xuezi Dan of Covington & Burling provide a breakdown of practical obstacles that AIGC providers could face according to the proposed regulations. This includes the thorny requirements (with good reason) for obtaining training data — the data must not infringe intellectual property, and data containing personal information must be obtained with consent from its subject or in a way that otherwise complies with the law.
I also participated in this forum and chimed in with some thoughts of my own. For instance, I believe that the solidification of AIGC regulations could be the push that China’s lagging SaaS industry3 has been waiting for.
King & Wood Mallesons: A Legal Perspective (Chinese)
Titled “Do Not Go Gentle Into That Good Night: Thoughts on the Generative AI Draft Regulations,” leading Asia law firm King & Wood Mallesons (KWM) lays out a nuanced analysis of China’s legal framework for AI-related services.
Calling the draft regulations “a nimble attempt at regulation in prompt response to the risks and impacts currently presented by generative AI,” the piece analyzes the regulations in great detail and discusses how they fit into China’s existing tech regulatory framework. It also compares China’s regulatory approach to other global pieces of tech policy, such as the EU’s Artificial Intelligence Act and the US’s Algorithmic Accountability Act of 2022.
Thoughts on Chinese AI and LLMs
ChatGPT and China: How to think about Large Language Models and the generative AI race (English)
Written by Paul Triolo for The China Project, this article examines the state of AI in China and explains how its growth will not follow the same path as AI tech in countries such as the US.
This piece expands its scope beyond the draft regulations to examine what China’s AIGC development may look like in the future. This discussion includes more abstract issues like political alignment, ethics, and governance, alongside very tangible obstacles such as access to advanced GPUs for training AI models.
Recent Trends in China’s Large Language Model Landscape (English)
This recent study by Jeffrey Ding & Jenny W. Xiao examines 26 Chinese LLMs (large language models) developed by Baidu, Inspur, Tsinghua University, Alibaba, and others. At about a dozen pages, it’s a relatively quick read that packs a lot of information. It also pairs nicely with the above article to provide a good picture of China’s AI landscape.
I highly recommend this report to anyone seeking a more nuanced understanding of China’s level of AI development. Some key points include:
How China’s LLMs measure up to those in the West (they are, in fact, “not far behind the state of the art”).
The prominence of AI governance and ethics in the development of these models.
The role of government sponsorship in AI development.
China’s increasing use of domestic hardware and software, which coincides with an increasing interest and pride in domestic technological innovation.
Bonus!
Growth of AI Through a Cloud Lens (English)
This last recommendation has nothing to do with Chinese AI (not directly, at least), but that should not deter you from reading it.
Written by HashiCorp founder Mitchell Hashimoto, this article frames current trends in AI as a “platform shift” mirroring the move to “cloud native” that occurred nearly two decades ago, when AWS’s release of S3 and EC2 paved the way for a cloud-first mindset that transformed how applications were developed, deployed, and scaled.
While AI will certainly follow its own growth path, the key steps of cloud’s growth as a mainstream paradigm can act as touchstones. This includes cloud’s early promises of immediate (and sometimes ephemeral) value, as well as its eventual evolution from an often impractical solution to one that engineers and businesses alike take for granted.
(Thanks to Kevin Xu for finding this article and Jordan Schneider for bringing it to my attention.)
For full English translations of the document, feel free to consult the separate versions published by DigiChina and China Law Translate.
See the “Deep synthesis technology” section of last month’s article on China’s generative AI landscape for more details.
In terms of cloud-based services, SaaS has consistently underperformed in China compared to global markets (see my summary of CAICT’s 2022 cloud computing white paper from last year for more background). With the rise of cloud-based AIGC services (as well as the beginning of a more consolidated digitalization policy approach from the central government, beginning with the Digital China plan), this could be about to change.
Many thanks!!