Understanding Comprehensive Legislation on Artificial Intelligence

On May 15, 2026, the General Office of the State Council issued the “2026 Legislative Work Plan,” which clearly proposed to accelerate the comprehensive legislation for the healthy development of artificial intelligence. This news has sparked widespread attention in various circles.

Discussions about AI legislation have always had differing opinions. Some believe that the current legislative need is not urgent and worry that legislation may hinder development. Others argue that AI has already been embedded in various scenarios, making it difficult to have a single law for comprehensive management.

Zhou Hui, Executive Deputy Secretary-General of the Internet and Information Law Research Association of the China Law Society and Deputy Director of the Internet and Information Law Office of the Chinese Academy of Social Sciences, believes that the risks of artificial intelligence are diffuse and cumulative. The sooner a predictable institutional framework is established, the lower the governance costs will be. Good legislation, in essence, promotes development.

“It is not simply a matter of whether to regulate, but rather how to regulate effectively,” he pointed out.

Regarding the understanding of “comprehensive,” he proposed that it does not mean creating an all-encompassing, detailed AI code, but rather a “basic law” or “general law” in the field of artificial intelligence, reserving institutional interfaces for subsequent supporting policies.

Currently, China has issued policy documents on generative AI and anthropomorphized AI. Zhou Hui believes that unified legislation on artificial intelligence will enhance the stability of rules, improve governance coordination, reduce the institutional costs brought by fragmented governance, address issues that low-tier rules cannot solve, and enhance international competitiveness.

The Need for Effective Regulation

Q: There has always been a debate about balancing development and regulation in AI legislation. Some argue that the current legislative need is not urgent. What is your view?

Zhou Hui: I believe that AI legislation is not a binary choice between “development” and “regulation.” The real discussion should focus on how to provide a more stable institutional environment for AI development through scientific, moderate, and effective rules.

We cannot apply the old experience of “develop first, regulate later” from the early internet era. As a general-purpose technology, AI carries diffuse and cumulative risks that span the entire process from design, training, to deployment. If we wait for risks to manifest on a large scale or even lead to systemic crises before regulating reactively, the governance costs will be extremely high and may accumulate irreversible risks.

Of course, emphasizing the necessity of legislation does not mean adopting simple strong regulation, nor does it mean stifling AI technology. Good legislation should promote development. It should clarify what behaviors are permissible, what are not, and which scenarios require higher safety standards, thereby reducing uncertainty and trial-and-error costs for enterprises.

Thus, I do not agree with the assessment that “the current legislative need is not urgent.” Precisely because AI is developing rapidly, has a wide impact, and entails complex risks, we need to establish a predictable, coordinated, and updatable institutional framework through legislation to promote the healthy development of AI on a legal basis.

Q: The “2026 Legislative Work Plan” proposes to accelerate comprehensive legislation for the healthy development of AI. How should we understand “comprehensive”? What does it mean?

Zhou Hui: To understand “comprehensive,” we must first clarify a misconception: we are not aiming to create an all-encompassing AI code. Comprehensive AI legislation should be positioned as a “basic law” or “general law” in the field of AI, responding to fundamental and global issues in AI development and governance.

The connotation of this “comprehensive” is very rich. In terms of content, AI governance cannot only focus on algorithms or data; it must also coordinate models, computing power, applications, responsibilities, regulation, and promotion policies, forming an interconnected institutional system.

In terms of scope, AI risks do not only occur at the application end; many issues arise during the design, training, and deployment phases. Therefore, legislation must cover the entire process from research and development to application, from launch to withdrawal, rather than just remedying problems after they occur.

In terms of functional positioning, AI legislation cannot be merely a “regulatory law” or just a “promotional law.” It should simultaneously fulfill multiple functions: promoting development, preventing risks, safeguarding rights, and regulating power.

This also means that comprehensive AI legislation should play a foundational and leading role. It should establish basic principles, basic systems, basic rights and responsibilities, and regulatory frameworks while maintaining sufficient flexibility to reserve broad institutional interfaces for subsequent supporting regulations, industry standards, local pilot projects, and specific scenario governance.

Embracing Diverse Scenarios in Legislation

Q: As AI technology is widely embedded in social work and daily life, scenarios become increasingly complex, and risks may differ across fields and scenarios. How will legislation accommodate these different scenarios?

Zhou Hui: AI legislation must accommodate different scenarios, and the key is not to apply a rigid set of rules to all problems but to establish a system structure of “unified baseline, categorized governance, and dynamic adjustment.”

First, we need to establish unified basic rules, such as safety, transparency, accountability, protection of personal rights, human oversight, and risk assessment, which should become the common baseline for AI research and application.

Second, we must acknowledge the risk differences between various scenarios. Fields like healthcare, finance, transportation, education, government, and judiciary have different risk natures and consequences, and a one-size-fits-all standard cannot be simply applied. For example, government and judicial scenarios involve the operation of public power and should emphasize procedural fairness, transparency, and final accountability; healthcare scenarios involve life and health and should prioritize safety verification and accountability tracing; open-source models require a special balance between safety obligations and innovation space.

Lastly, we need to maintain flexibility through institutional design. AI technology evolves rapidly, and legislation cannot predict all scenarios at once. Therefore, mechanisms such as negative lists, regulatory sandboxes, local pilot projects, and dynamic assessments can allow rules to be continuously adjusted with technological advancements and application changes.

Thus, comprehensive legislation does not exclude scenario-based governance. On the contrary, it provides a unified framework and institutional interfaces for governance in different scenarios, enabling differentiated and refined governance under common rules.

Q: In recent years, numerous policy documents related to AI have been issued. What does higher-level legislation mean, and what impacts will it bring?

Zhou Hui: The departmental regulations and local policies issued in recent years have indeed played a timely guiding role, but as the industry deepens, their limitations have become increasingly evident. The main pain point lies in the lower effectiveness level, lack of coordination, and the tendency to lead to overlapping and vacant regulations, resulting in a “nine dragons managing water” phenomenon. Relying solely on fragmented policies makes it difficult to provide enterprises with stable and authoritative compliance expectations.

I believe that higher-level legislation means that AI governance will rise from policy guidance, departmental regulations, and local explorations to a unified legal institutional arrangement at the national level, which will at least bring the following impacts:

1. Enhancing rule stability. Enterprises can better assess future institutional directions, forming more stable expectations for research, investment, and compliance.
2. Improving governance coordination. By unifying basic principles, regulatory systems, rights and obligations, and responsibility rules through law, we can reduce the institutional costs brought by fragmented governance.
3. Addressing issues that lower-tier rules cannot solve. For instance, civil liability allocation, liability exemptions, tax support, licensing access, and foundational rights and obligations arrangements often require legal basis.
4. Enhancing international rule competitiveness. AI governance has become an important part of global competition. Through high-level legislation, we can more clearly express our basic positions in AI development and governance, which also helps enhance our international discourse power.

Q: Are there currently legal regulations that do not match the development of AI, and how will legislation address these gaps?

Zhou Hui: There are indeed mismatches, and this mismatch is not just an issue of individual provisions but a deeper systemic problem.

1. Existing rules do not sufficiently cover core elements of AI. Current systems focus more on issues like data, cybersecurity, personal information protection, and algorithm recommendations, but the development of AI also involves models, computing power, foundational large models, open-source models, model parameters, and the legal use of training data, which lack systematic rules.
2. Traditional legal rules struggle to adequately respond to new legal relationships brought about by AI. For example, where are the boundaries for using copyright-protected data in model training? How is the legal status of AI-generated content determined? How is liability allocated when foundational models have defects that affect downstream applications? How can victims seek redress when algorithmic black boxes make it difficult to provide evidence? These issues require new institutional responses.
3. Some old rules may hinder innovation. For instance, copyright rules, data circulation rules, traffic management rules, and product liability rules, if not adapted to AI scenarios, may prevent enterprises from innovating, localities from piloting, and regulatory bodies from grasping boundaries.

Therefore, to address these shortcomings, legislation must innovate paradigms, specifically focusing on the following aspects:

1. Establishing an incentive and protection system covering all elements, including data, algorithms, models, and computing power, such as exploring statutory licenses for model training and special subsidies for computing power construction, providing foundational rules for AI development.
2. Creating a responsibility system divided by subject, breaking the single responsibility framework, accurately allocating obligations based on the technical control and profit acquisition of developers, providers, and users, while distinguishing the obligations and responsibilities of different entities such as platforms, computing service providers, government departments, and open-source contributors.
3. Establishing a full lifecycle governance mechanism, moving risk prevention to the design, development, training, and deployment stages, rather than waiting until damage occurs to assign accountability.
4. Establishing dynamic adaptation and suspension mechanisms, defining the pivotal role of the national AI regulatory authority, and allowing the legal system to proactively shape the technological environment through legislative evaluations, regulatory sandboxes, and other interfaces.

Ultimately, the goal of AI legislation is not merely to address the question of whether to regulate, but to further tackle how to regulate effectively, that is, how to create a systematic, stable, open, and updatable legal framework that allows AI to unleash greater development potential on a safe and controllable basis.