Regulation, Law & Liability
Early and stringent AI regulation — including the Generative AI Governance Regulation of 2023 — is state-directed, offering legal certainty within a framework of limited openness.
Where does China stand, and where might it break?
In August 2023, China enacted one of the world's first binding frameworks for generative AI with the Interim Measures for the Administration of Generative Artificial Intelligence Services, led by the Cyberspace Administration of China (CAC). Providers must register their algorithms, undergo security assessments, and align content with state requirements, including the prescribed value system and information controls. Mandatory labelling rules for AI-generated content entered into force in September 2025. These rules sit within an already dense legal framework: the Cybersecurity Law of 2017, the Data Security Law of 2021, and the Personal Information Protection Law (PIPL) of 2021, which draws on European standards whilst leaving broad scope for state access to data. Further regulations govern recommendation algorithms and "deep synthesis" — AI-generated images, audio, and video. A comprehensive, unified AI law is in preparation but has not yet entered into force; draft versions are being discussed in specialist circles. China's regulatory approach is therefore early, detailed, and firmly enforced, yet centrally directed and in the service of state control objectives. This means legal certainty combined with limited openness.
When AI systems make decisions, generate content, and take over tasks, new questions of liability, accountability, and control arise. China's response is early, state-directed regulation. It provides legal certainty, but subordinates the direction of innovation to the interests of the state. This gives rise to a dual effect. Clear rules reduce uncertainty for providers and accelerate adoption. At the same time, the state's control over content and data limits diversity, error correction, and independent oversight. For the question of liability, this means that responsibility rests primarily with registered providers and, ultimately, with the state as rule-setter and data custodian, rather than with independent courts or affected individuals. Whether this model — combining early regulation with state direction — can sustain both innovation and trust, or whether it sacrifices openness and legal protection, will shape the governance of the AI economy through to 2040.
Paths for China
China is consolidating its scattered regulations into a comprehensive AI law with tiered risk categories and clear liability rules for providers, operators, and users. This matters because the current patchwork of rules leaves legal uncertainty at the margins, and a unified law would assign responsibility unambiguously. Draft legislation is already under discussion. The trade-off: a rigid risk framework may constrain innovation, and in a state-directed system without independent courts, liability remains of limited enforceability.
Trade-off: Legal certainty versus flexibility and independent oversight
Feasibility: Effective only with robust, independent enforcement
China is expanding and strictly enforcing its existing registration and labelling requirements for algorithms and AI-generated content, with the aim of curbing manipulation, disinformation, and misuse. This addresses the core of the structural break, as transparency over AI-generated content builds trust — and the labelling obligation that entered into force in 2025 already points in this direction. The trade-off is that the same instruments can serve information control, and that stringent requirements may burden smaller providers and open research.
Trade-off: Trust and protection versus openness and plurality of opinion
Feasibility: Impact depends on whether control or protection prevails
Building on the PIPL, China is strengthening individual data rights and establishing effective channels for complaints and redress against flawed or harmful AI decisions. This matters because trust in AI depends on enforceable rights for those affected, not merely on obligations placed on providers. The cost lies in the tension with the state's far-reaching data access. As long as the state acts as central data custodian, individual rights vis-à-vis public authorities remain limited, and legal protection operates primarily against private providers.
Trade-off: Individual legal protection versus state data sovereignty
Feasibility: Effective primarily against private actors; scarcely so against public ones
Trigger-based phases
Trigger: Binding rules for generative AI, algorithms, and deep synthesis enter into force and are enforced.
China establishes legal certainty for providers at an early stage, but ties it to registration and content requirements. Labelling obligations for AI-generated content are introduced, and state oversight is centralised.
Trigger: The scattered regulations are consolidated into a comprehensive AI law with risk categories and liability rules.
China harmonises its regulatory framework and allocates responsibility more clearly. The balance between legal certainty, innovation, and state control becomes the central question.
Trigger: The Chinese model of early regulation and state direction is established and begins to influence international standards.
China is co-authoring global AI standards and exporting its governance model. Whether it sustains innovation and trust, or sacrifices openness and legal protection, will be decided by 2040.
Real-world pilot programmes
Status: In force (one of the world's first binding AI regulations)
The interim measures issued by the CAC in August 2023 govern generative AI services through registration, safety, and content obligations. They are the test case for whether early, state-directed rules can deliver both legal certainty and innovation.
Status: In force since September 2025 (mandatory labelling)
Since September 2025, AI-generated content must be visibly labelled and marked in metadata. The rule serves as a model case for how transparency around AI content can build trust while simultaneously serving the purposes of information control.
)
Status: In force since 2021 (data protection under state data sovereignty)
The 2021 PIPL introduced data-protection rules modelled on European standards, while state-run data exchanges trade data as a factor of production. Together, they illustrate China's balance between individual data protection and far-reaching state access to data.
What we are watching
The measure tracks how clearly responsibility and liability are regulated. The threshold matters because only a unified law can deliver legal certainty beyond the gaps left by today's patchwork. Its entry into force will show how China assigns liability in an AI economy and how it balances innovation against control.
The indicator tracks how transparently AI-generated content is identified. The threshold matters because transparency builds trust and curbs manipulation. If labelling is applied consistently, the risk of disinformation falls. If it remains patchy or serves primarily the purposes of control, the balance between protection and direction shifts.
The measure tracks how effectively those affected can challenge AI decisions. The threshold matters because trust in AI also depends on enforceable rights. If effective channels for redress are established, trust increases. If they remain weak in relation to public authorities, legal protection stays one-sided.