China's AI rules and the EU AI Act: a side-by-side comparison (2026)

ai Oct 29, 2026 5 min read Caybles Research Desk

TL;DR

China and the EU have built the world's two most consequential AI regimes, on opposite architectures. China governs through vertical, application-specific instruments — algorithms, deep synthesis, generative AI services, content labelling, and now draft rules for human-like AI — enforced by the CAC and sectoral ministries through filings and security assessments. The EU AI Act governs horizontally by risk tier, fully applicable since 2 August 2026, with penalties reaching 7 percent of global turnover. A company building for both markets faces overlapping but structurally different obligations, and a comprehensive Chinese AI law remains drafted rather than enacted.

Two AI regulatory systems now shape how most of the world's technology companies build and sell, and they could hardly be more different in design. Understanding the difference — vertical application rules in Beijing, horizontal risk tiers in Brussels — is more useful than any attempt to rank the two by strictness. This comparison sets out how each system actually works, side by side, for teams that have to satisfy both.

The comparison at a glance

Dimension	China	European Union
Core instruments	Generative AI measures (2023); content-labelling measures (2025); draft human-like AI rules (2025); algorithm and deep-synthesis rules (2022–23)	AI Act (2024), fully applicable 2 Aug 2026
Architecture	Vertical: rules attach to specific applications and services	Horizontal: rules attach to risk tiers of the use case
Model-level duties	Security assessments and filings for public-facing services	GPAI transparency and documentation; systemic-risk tier
Content transparency	Mandatory labels and watermarks on AI-generated content	Transparency duties for synthetic content and chat interfaces
Enforcement	CAC with sectoral ministries (MIIT, MPS and others)	Member-state authorities plus the EU AI Office
Penalties	Service rectification, fines, app-store removal	Up to 7% of global turnover or €35M for prohibited practices
Comprehensive law	Drafted, not enacted; on the State Council legislative plan	Enacted; implementation under way

China: rules that attach to applications

Beijing has never passed a single omnibus AI statute. Instead, the Cyberspace Administration of China and partner ministries have issued a succession of targeted instruments, each attaching to a defined class of service. Recommendation algorithms came first, in 2022, followed by deep-synthesis rules in early 2023 and interim measures for generative AI services in August 2023. The pattern continued with binding content-labelling measures effective from September 2025 and draft rules on human-like interactive AI — chatbots, companions and virtual humans — issued for comment in December 2025, which Mayer Brown's analysis describes as introducing technical safeguards and psychological-safety obligations for providers of emotionally interactive services.

The machinery underneath is administrative: public-facing generative services undergo security assessments and complete filings before launch, and new AI ethics-review rules — including a ten-province pilot running from June to November 2026, tracked by Concordia AI — push registered ethics committees into corporate governance. A comprehensive AI Law has appeared in the State Council's legislative work plans for several consecutive years without advancing to enactment, a status Bird & Bird's regulatory tracker still reports, with 2026 attention centred on a planned comprehensive law, an updated Cybersecurity Law and the anthropomorphic-AI rules according to China Briefing.

What this architecture means in practice: obligations arrive earlier in the product lifecycle (a filing precedes a launch), and they are legible service by service — but they shift as each new application class gets its own instrument. The Caybles policy tracker indexes each of these documents, from CAC, MIIT, NDRC and MOFCOM, as it is published.

The EU AI Act: rules that attach to risk

Brussels chose the opposite axis. The AI Act does not ask what a system does but how risky its use case is, grading deployments from minimal to unacceptable and attaching duties accordingly. Prohibitions on unacceptable-risk practices took effect in February 2025; obligations for general-purpose models followed in August 2025; and the Act became fully applicable on 2 August 2026, per the European Commission, with penalties for the gravest violations reaching 7 percent of global turnover or €35 million.

Two features distinguish the European approach. Extraterritorial reach: the Act applies to AI systems whose outputs are used in the Union, regardless of where the provider sits — which is why Chinese models face the AI Act too, as our DeepSeek analysis discusses. And contested momentum: Brussels has debated targeted delays to parts of the high-risk rulebook through its digital-omnibus package, a signal that while the architecture is settled, implementation timing remains political.

For all the architectural difference, convergence is real on three points. Both mandate transparency toward users — China through labelling and watermarking duties, the EU through disclosure and synthetic-content markings. Both single out interactive and companion-style AI for special attention: China's draft human-like AI measures on one side, the EU's deployer duties around emotionally close use cases on the other. And both have discovered the same enforcement truth, that obligations without administrative capacity are paper — which is why China's ethics-committee pilot and the EU's authority staffing drawdowns matter as much as any article of either framework.

Where companies get caught

The bind is most visible for a product that must satisfy both regimes at once. A conversational assistant offered to consumers in both markets triggers China's filing-and-assessment pathway and the EU's transparency and deployer duties, under different vocabularies and different enforcers. Compliance teams handle this by mapping obligations per market rather than seeking a single global posture: the CAC-and-ministries explainer covers who actually holds the pen in Beijing, and the Act's official texts remain the operative reference for Brussels.

Frequently asked questions

Which regime is stricter?

The question assumes a single axis. China's regime is stricter at market entry — the filing and security-assessment pathway gates launch — while the EU's is stricter at penalty: 7 percent of global turnover has no Chinese equivalent. They are strict about different moments of a product's life.

Does the EU AI Act apply to Chinese AI models?

Yes. The Act reaches systems whose outputs are used in the Union, wherever the provider is established, and its GPAI obligations cover models placed on the EU market.

Has China passed its comprehensive AI law?

Not yet. The AI Law has sat on the State Council's legislative plan for several years without a published draft text; the operative rules remain the application-specific measures.

Do Chinese AI labelling rules apply to foreign companies?

The labelling measures attach to generative services operated within China. A foreign provider serving Chinese users through a domestic operation is within scope; enforcement against offshore-only providers remains a developing question.

Sources

Not advice. This article is published for general information and analysis only. It is not legal, regulatory, investment or professional advice, and it should not be relied upon as such — please do your own due diligence and consult qualified counsel for your specific situation. Regulatory positions described here may have changed since publication.