China's AI Regulation Blueprint: Balancing Innovation and Control

China's Vision: From National Priority to Global AI Leader

In October 2017, the People's Republic of China's (PRC) State Council, the country's highest administrative authority, declared artificial intelligence a "national priority," stating the country's ambition to become "the world's premier artificial intelligence innovation center" by 2030. The stakes are high because Chinese AI has medical, industrial, and transportation applications, particularly in self-driving cars, and could be a significant driver of the country's growth (China's Race to Become an AI Superpower | the Nation, 2023).

China also has a large pool of skilled workers; nearly 1.4 million engineers qualify yearly, six times as many as in the United States, with at least one-third specializing in AI. According to the Japanese business daily Nikkei Asia, "China is the undisputed champion in artificial intelligence research papers, far surpassing the United States in both quantity and quality." Tencent, Alibaba, and Huawei are now among the top ten global producers of this kind of output.

China has also emerged as an early adopter and testing ground for national AI regulations. Since 2017, China has issued binding regulations for recommendation algorithms, deep synthesis content or deepfakes, and generative AI systems. These regulations establish requirements for algorithms and AI models' development, validation, and deployment.

China's AI regulations kicked off in 2017 with the release of the New Generation AI Development Plan, which primarily focuses on promoting AI development. It also outlines a high-level timeline for developing AI governance regulations through 2030.?

AI Governance: From Early Moves to Concrete Measures

China has issued nine AI-related rules or regulations since then. The three principal regulations that require attention and analysis are summarized below:

1) Algorithmic Recommendation Management in Internet Information Services. Known as the Regulation of "Recommendation Algorithms"

Date of issuance: December 31, 2021?

It is the first centrally binding algorithm regulation, motivated by fears regarding algorithms controlling how news and content are distributed online. The regulation includes many provisions for content control and worker protections impacted by algorithms. It also established the "algorithm registry" that will be used in future regulations.

(See The translated version of the regulation.) ?

The Chinese Communist Party's "CCP" initial motivation for the regulation was a concern about the role of algorithms in online information dissemination and the fear of losing control of mass AI-generated content.

The algorithm registry (literally "algorithm filing system" in Chinese) was created due to the recommendation algorithm regulation. The registry is an online algorithm database.?

Developers of these algorithms must submit information on how they are trained and deployed, including which datasets are used to train the algorithm. In addition, they must submit an "algorithm security or safety self-assessment report." Following the successful registration of an algorithm, a limited filing version is made public. The following issued regulations on deep synthesis content and generative AI, required the developers to register their algorithms.

2) Rules for the Management of Deep Synthesis Internet Information Services.

Date of Issuance: November 25, 2022.

At the same time that the CCP was concerned about the recommendation algorithms (2017-2019), it identified deepfakes as a significant threat to its information environment and began regulating them. Tencent, the giant technology company, introduced and popularized the term "deep synthesis" to describe the synthetic generation of content during the policy development process, replacing the politically charged "deepfakes" with a more innocuous-sounding technical term.?

The new term eventually gained traction and was incorporated into the final rule. The deep synthesis regulation was broadened to include algorithms for generating or altering online content such as voice, text, image, and video.?

Deep synthesis content must adhere to information controls and be labeled as synthetically generated, and providers must take steps to prevent misuse.?

The regulation includes several ambiguous censorship requirements, such as requiring deep synthesis content to "adhere to the correct political direction," not to "disturb economic and social order," and not to be used to generate fake news. When such content "might confuse or mislead the public," a "conspicuous label in a reasonable position" must be included to alert the public that it was generated synthetically.?

The regulation also includes anti-misuse provisions, such as requiring deep synthesis users to register with their real names and platforms to prompt users to obtain the consent of anyone whose personal information is being edited. Finally, deep synthesis providers must register with the algorithm registry.?

The deep synthesis regulation was finalized on November 25, 2022, five days before OpenAI's ChatGPT was released.?

3) Management Measures for Generative Artificial Intelligence Services

Draft Issued: April 11, 2023 (Draft for comment, not enacted yet)

The regulation, drafted by the Cyberspace Administration of China (CAC) in response to the surge in popularity of AI chatbots such as ChatGPT, covers nearly the same ground as the deep synthesis regulation but focuses on text generation and training data. Providers must ensure that the training data and content generated are "true and accurate."?

China's regulatory apparatus initially appeared well-prepared for the wave of generative AI applications that would follow ChatGPT. Most forms of generative AI used to create or edit images, videos, voice, and text were technically included in the deep synthesis regulation.?

However, CAC officials deemed the newly enacted deep synthesis regulation insufficient. Deepfakes were the primary concern behind the deep synthesis measures, and their labeling requirements might make sense for visual or audio deepfakes. Still, more is needed to address new concerns about text generated by large language models (LLMs) or the technology's increasingly general-purpose nature.?

Furthermore, the original regulation only covered deep synthesis services delivered over the Internet, leaving a regulatory gap for generative AI services delivered offline. As a result, Chinese regulators and policymakers rushed to draft a new regulation covering nearly the same set of AI applications but with an updated set of concerns.?

The draft regulation of the new generative AI regulation was issued for public comment in April 2023. Many boilerplate content mandates ("embody Core Socialist Values") were reinforced in the draft, and providers were required to submit a filing to the existing algorithm registry. It also included several new requirements for training data and generated content that providers may find extremely difficult to meet. According to the draft, the providers must ensure the "truth, accuracy, objectivity, and diversity" of their training data. This is a potentially impossible standard for LLMs trained on massive troves of text and images scraped from millions of websites.?

The draft requires that training data not infringe on intellectual property rights. The regulation requires that generative AI not be discriminatory based on race or gender and that generated content be "true and accurate," posing a technical challenge for LLMs prone to "hallucinating" inaccurate or baseless claims in their outputs.?

These stringent requirements for generative AI systems have sparked a heated debate over the draft regulation. When writing this article, Chinese academics, businesses, and policymakers are debating how to maintain adequate content controls without suffocating China's new generative AI industry.?

The Balancing Act: Control vs. AI Leadership

While promoting Chinese leadership in AI, the different regulations impose controls to serve the political agenda of the Chinese Communist Party. This includes restrictions on content and government disclosures to shape public discourse and maintain social stability.?

China's regulatory approach will shape international tech development as a dominant global AI superpower. Chinese restrictions will impact the ability of Chinese tech companies to grow and dominate AI industries worldwide. Understanding the regulations helps countries and companies engage with China's AI ecosystem.

Drawing Lessons from China's AI Regulations

While democracies may disagree with China's goals, they can learn from its regulatory experiments. China's technical requirements and oversight mechanisms highlight both opportunities and risks for regulating the societal impacts of AI. A pragmatic review of China's AI governance will help countries and policymakers develop informed national AI policies.

In AI, where technology leaps ahead of governance, China's regulations offer early glimpses of what's possible. Globally, AI regulation is contested, fragmented, and evolving.?

What are the most important takeaways from China's moves to regulate AI? What can democracies learn from effectively governing this transformative technology? Please feel free to leave your thoughts in the comments.

References:

• China's Race to Become an AI Superpower | The Nation. (2023, April 10).
• Sheehan, M., 2023. China's AI Regulations and How They Get Made. United States of America. Retrieved from?https://policycommons.net/artifacts/4465013/chinas-ai-regulations-and-how-they-get-made/5262348/?on August 19, 2023. CID: 20.500.12592/8rb2bt