Last reviewed: April 9, 2026
Jurisdictions covered: China (primary), EU and Korea (comparison)
Reading time: 12 minutes
China AI Ethics: From Voluntary Principles to Mandatory Review in Five Years
In September 2021, China’s Ministry of Science and Technology published voluntary AI ethics norms. Six principles. No enforcement mechanism. No penalties.
In April 2026, ten government agencies jointly issued mandatory ethics review measures. The same six principles. But now with a three-tier review process, a high-risk activities list, and binding obligations for every organization developing or deploying AI.
The principles did not change. The enforcement mechanism did.
This pattern — voluntary guidelines hardening into binding rules — is the single most useful lens for predicting where Chinese AI regulation goes next. The algorithm recommendation rules followed this path. The deep synthesis rules followed this path. The generative AI measures followed this path. And now, AI ethics review has crossed the same threshold.
This article traces the trajectory from 2021 to 2026, explains what the new mandatory framework requires, and maps where voluntary guidelines suggest binding rules will appear next. For the full regulatory picture, see our China AI Regulation 2026 guide.
Key Takeaways
- Voluntary → mandatory in 5 years. The 2021 MOST Ethics Norms became the 2026 10-Agency Ethics Review Measures. The pattern is consistent: China publishes voluntary guidelines, tests industry reaction, then codifies binding rules.
- 10 agencies, 6 review areas. Led by MIIT, the April 2026 measures require ethics review covering: human welfare, fairness, controllability and trustworthiness, transparency and explainability, accountability and traceability, privacy protection.
- Three-tier review process. Internal ethics committee → ethics review service center → government expert review. Higher-risk activities escalate to higher tiers.
- “High-risk activities” trigger enhanced review: AI that mobilizes public opinion, makes autonomous decisions, or enables deep human-machine integration.
- This is not decorative. The measures apply to real R&D and deployment decisions. Organizations with AI projects in China must assess whether they fall in scope.
The 2021 MOST Ethics Norms: Where It Started
The New Generation AI Ethics Norms (September 2021) established six foundational principles. These were voluntary — no enforcement mechanism, no penalties.
The 6 principles:
1. Enhancing human welfare. AI should serve human interests, improve quality of life, and promote sustainable development.
2. Promoting fairness and justice. AI design and use should avoid discrimination and bias.
3. Protecting privacy and security. AI should respect personal information, data security, and national security.
4. Ensuring controllability and trustworthiness. AI systems should remain under human control with predictable, reliable behavior.
5. Strengthening accountability. Clear responsibility chains from development through deployment.
6. Improving ethical literacy. AI practitioners should receive ethics education and training.
These principles were not enforceable. MOST published them alongside a set of “ethical requirements” for AI management, R&D, supply, and use — but without regulatory teeth. Industry response was mixed: major companies (Baidu, Alibaba, Tencent) appointed internal ethics officers; smaller companies largely ignored the guidelines.
Source: CSET Georgetown translation, October 2021; MERICS analysis, Rebecca Arcesati, June 2021.
How Ethics Guidelines Map to Existing Binding Rules
The ethics framework does not exist in isolation. Each principle connects to at least one binding regulation already in force:
| Ethics Principle | Binding Rule That Implements It | Article |
|---|---|---|
| Fairness | PIPL Art. 24 — non-discrimination in automated decisions | Prohibits “unreasonable differential treatment” |
| Fairness | Algorithm Recommendation Rules — price discrimination ban | Art. 12 |
| Transparency | Deep Synthesis Rules — content labeling | Art. 16-17 |
| Transparency | GenAI Measures — algorithm filing disclosure | Art. 17 |
| Privacy | PIPL — consent, cross-border transfers, automated decisions | Multiple articles |
| Controllability | GenAI Measures — content review systems, human oversight | Art. 4-5 |
| Accountability | CAC filing — provider identity and responsibility | Filing requirements |
| Human welfare | Facial Recognition Measures — alternatives required, minors protection | Art. 7, 10 |
Our view: The ethics review framework fills the gap between principles and enforcement. Before April 2026, a company could comply with every binding rule and still have no process for assessing whether its AI was “fair” or “controllable” in a broader sense. The ethics review measures require that assessment — and they require it to be documented, reviewable, and potentially escalated to government experts.
What These Guidelines Signal for the Future
If the pattern holds, three currently voluntary frameworks will become binding:
1. AI Safety Governance Framework v2.0 (September 2024). This TC260 framework introduces a risk-management methodology with three risk categories (inherent technology risk, application risk, societal risk). If it follows the ethics path, binding risk management requirements could appear by 2026-2027 — potentially as part of the comprehensive national AI law under legislative research.
2. AI agent governance. TC260 issued draft AI agent security guidelines in April 2026. Agent-specific binding rules could follow within 12-18 months if the pattern holds.
3. Compute governance. China has drafted provisions for regulating foundation models by compute, parameters, or scale. A voluntary governance framework is likely before binding compute thresholds. The EU’s 10^25 FLOP threshold provides a reference point.
Disclaimer: This content is for informational purposes only and does not constitute legal advice. Chinese AI regulation changes frequently. Reg Intel is not a law firm and does not provide legal services.
Last verified: April 9, 2026
Compare: EU vs China
For the global keystone comparison across twelve dimensions — algorithm filing vs conformity assessment, content moderation conflicts, asymmetric extraterritoriality, enforcement philosophy, and a five-step dual-market compliance baseline — see EU vs China AI Regulation: Two Systems, Two Philosophies (2026).