← Back to feed

On April 3, 2026, China’s Ministry of Industry and Information Technology, together with nine other agencies including the National Development and Reform Commission and the Chinese Academy of Sciences, issued the Administrative Measures for the Ethical Review and Services of Artificial Intelligence Science and Technology (Trial). The rules are effective immediately.

The scope is sweeping. Every organisation engaged in AI science and technology activities — companies, universities, research institutions, hospitals — must establish an internal AI Science and Technology Ethics Committee. Projects cannot proceed until that committee approves them. If an organisation lacks the capacity to conduct its own review, it must outsource to an approved external “ethics review service centre.”

Three-Tier Enforcement Architecture

The system is structured in three layers:

1. Internal ethics committees — mandatory within every qualifying organisation; primary responsibility sits here
2. External service centres — licensed bodies that conduct reviews for organisations that cannot run their own
3. Government-led expert review — mandatory for projects on a separately published “high-risk activities list”

High-risk triggers include AI systems capable of “influencing social discourse,” models with “highly autonomous decision-making,” and applications involving “deep human-machine integration.” These categories require external expert panel sign-off before development can commence, not merely after deployment.

What Reviewers Must Assess

The measures define specific review criteria: training data selection methodology, algorithm and model design rationality, bias and discrimination prevention measures, robustness in open and adversarial environments, and whether users retain the ability to control and intervene in AI systems. Review bodies must issue decisions within 30 days of receiving a complete application; projects on the high-risk list face a narrower 36-hour window under emergency procedures.

Approval is not permanent. Projects already underway are subject to ongoing monitoring, and review bodies can suspend or terminate them if risk conditions change. Organisations must file annual reports through a national science and technology ethics management platform.

China vs. the West on AI Governance Speed

The contrast with Western approaches is structural. The US White House legislative framework issued in recent months represents proposals to Congress — aspirational, subject to legislative process, and explicitly pre-empting state-level rules. China’s administrative measures are co-issued by 10 agencies and enforceable on the date of publication. No parliamentary vote, no comment period.

The European AI Act, which went into force in 2024, shares the risk-tier architecture but took years of legislative process to reach that point. China has implemented a comparable pre-approval regime via ministerial decree in weeks.

Implications for Chinese AI Labs

The timing lands as China’s frontier AI labs — Z.AI, MiniMax, ByteDance, Alibaba — are operating at a release cadence that rivals Western counterparts. GLM-5.1 launched March 27. MiniMax M2.5 dropped February 12. Regulatory overhead for each new model release now includes mandatory pre-review, raising the institutional cost of iteration.

The prior 2023 system that the new measures replace was widely criticised for ambiguity and weak enforcement. The new framework replaces discretionary guidance with procedural obligations: committees, service centres, registration platforms, and a tiered review regime. Compliance gaps are now enforceable under China’s Network Security Law, Data Security Law, and Personal Information Protection Law.

For foreign companies operating AI development in China, the committee establishment requirement applies to domestic entities. Subsidiaries and joint ventures engaging in qualifying AI activities in-country appear to be in scope — a question that compliance teams at Alphabet, Microsoft, and others operating China-registered entities will need to resolve.