Your company sells AI in Europe and is being asked to do the same in China, or the reverse. Two regulators now demand evidence before your model reaches users, and they ask for different things. The EU AI Act asks where your system sits in a risk hierarchy and whether you can prove conformity. China asks what your service does to public content and whether you have filed it with the Cyberspace Administration of China. Companies that treat these as two versions of one regime build the wrong paperwork twice. This guide compares them where compliance actually diverges, and shows how to run both without doubling the work.
Who this guide is for
- General counsel and compliance leads at AI companies serving, or planning to serve, both the EU and China.
- Teams using an EU AI Act programme as the template for a China launch.
- Investors and boards assessing regulatory exposure across the two largest AI markets.
Stage: Explore. You are mapping obligations across jurisdictions before committing budget or product architecture.
Two design philosophies
The EU AI Act is a product-safety statute. It treats an AI system as a product placed on a single market, classifies it by risk, and requires the provider to demonstrate conformity before release, in the way that machinery or medical devices are certified. Enforcement runs through market surveillance authorities, CE marking and post-market monitoring. Its central question is: can you prove this system is safe enough for its risk class?
China’s AI rules are a service-governance regime. They treat an AI service as something that reaches the public and shapes what the public sees, and they regulate the operator through registration, security assessment and continuous content responsibility. The core instruments are the Algorithm Recommendation Provisions, the Deep Synthesis Provisions, the Interim Measures for Generative AI Services and the AI-Generated Content Labelling Measures, layered on top of the Cybersecurity Law, Data Security Law and Personal Information Protection Law (PIPL). Its central question is: who is accountable for what this service puts in front of Chinese users?
The difference in philosophy explains most of the practical differences. The EU regulates before the market; China regulates before the public. The EU wants a technical file; China wants a filing, a responsible domestic entity and a moderation operation.
Scope, reach and how high-risk is defined
| Dimension | EU AI Act | China |
|---|---|---|
| Legal form | Regulation (EU) 2024/1689, directly applicable | Several administrative measures plus three baseline statutes |
| Regulatory model | Risk tiers with conformity assessment | Service-type registration plus sector regulation |
| Territorial reach | Providers placing on the EU market or whose output is used in the EU, wherever established | Services provided to the public in China, which in practice requires a domestic applicant |
| Risk classification | Prohibited, high, limited and minimal risk, with listed use cases | No tiers; triggers are public opinion attributes, social mobilisation capability, deep synthesis and sector rules |
| Main pre-launch step | Conformity assessment, CE marking, database registration for high-risk | Algorithm filing, generative AI security assessment and filing, security grading |
| Transparency | Disclosure of AI interaction, synthetic content marking, deepfake labelling | Visible and metadata labels on AI-generated content, plus published algorithm filing information |
| Enforcement style | Market surveillance, fines, withdrawal from market | Administrative review, rectification orders, suspension, app removal |
| Penalty ceiling | Up to tens of millions of euros or a share of worldwide turnover | Statutory fines under CSL and PIPL plus operational sanctions |
The extraterritorial point deserves emphasis. The EU AI Act applies to a provider wherever it is established if the system is placed on the EU market or its output is used in the EU. China’s measures apply to services provided to the public within China, and because the applicant for a filing must be a domestic legal person, an offshore provider cannot satisfy the regime from abroad. In practice, the Chinese regime’s extraterritoriality is enforced at the point of access: app stores, cloud infrastructure and network connectivity.
High-risk is defined by list in Europe. The Act names product-safety systems in its first annex and use cases in areas such as biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration and justice in its second. If your system falls inside a listed use case, you are high-risk, and the provider performs the assessment. China has no equivalent list. Its de facto high-risk categories emerge from three sources: whether the service has public opinion attributes or social mobilisation capability, whether it is a deep synthesis service, and which sector regulator owns the use case. Medical AI is regulated as a medical device by the National Medical Products Administration; automotive AI by vehicle regulators; financial AI by financial regulators. A Chinese company therefore asks not “what tier am I” but “which regulator owns me and what does that regulator require”.
Transparency, content and training data
Both regimes impose labelling, but for different reasons and at different points.
Europe requires transparency so that users know they are dealing with AI and so that synthetic content can be detected. The Act requires disclosure when a person interacts with an AI system, machine-readable marking of synthetic content, and clear disclosure of deepfakes. China requires labelling so that the origin of content is traceable through the whole distribution chain. The Labelling Measures in force since 2025 require both an explicit label that users can see and an implicit label embedded in metadata, and they place duties on the content producer, the distribution platform and the app store.
Content moderation is where the systems diverge most. China imposes proactive obligations: filtering inputs and outputs, maintaining complaint channels, responding within defined periods and reporting to the Cyberspace Administration of China (CAC). The EU imposes transparency, fundamental rights impact assessment for certain public-sector uses, and incident reporting, but does not require the provider to police lawful speech. A moderation operation built for China will exceed what Europe requires; a compliance programme built only for Europe will fail in China.
On training data, neither regime grants a general licence to scrape. Europe requires providers of general-purpose models to maintain a copyright policy, respect text and data mining opt-outs and publish a summary of training content. China’s Interim Measures require lawful data sources, respect for intellectual property, consent where personal information is involved, and measures to improve data quality. PIPL adds separate consent for sensitive personal information and an impact assessment for high-risk processing. In both markets, the defensible position is documented provenance: where each dataset came from, under what licence, and how personal information was handled.
Conformity assessment versus filing
This is the operational difference that most often derails a shared compliance plan.
| Element | EU AI Act | China |
|---|---|---|
| Who performs the check | Provider self-assessment; notified body for listed product-safety systems | Regulator reviews a submission; no equivalent of a notified body |
| Core document | Technical documentation, risk management file, quality management system | Filing form, algorithm description, security self-assessment report, data and moderation description |
| Applicant | Provider or authorised representative in the EU | Domestic legal person operating the service |
| Output | Declaration of conformity, CE marking, database entry | Filing number, published basic algorithm information |
| Typical timing | Weeks for self-assessment, longer where a notified body is involved | Commonly around 30 working days for algorithm filing, often longer for generative AI in practice |
| Ongoing duties | Post-market monitoring, serious incident reporting, quality system maintenance | Periodic reporting, content moderation, inspections, re-filing after material change |
The consequence is sequencing. In Europe you can design, document, self-assess and launch. In China you must have the entity, the domain and website filing, the security grading and the filing submission accepted before a public launch, which pushes legal work to the front of the product timeline. Companies that start the Chinese process at launch discover that the filing is the launch.
Penalties and enforcement
Europe’s penalties are financial and graduated. Breaching the prohibitions carries the highest tier, measured in tens of millions of euros or a percentage of worldwide annual turnover, whichever is higher. Other obligations carry lower tiers, with further reductions for supplying incorrect information. Small companies face capped amounts, but the reputational and procurement consequences of a finding are often larger than the fine.
China’s penalties are financial and operational. PIPL allows fines up to RMB 50 million or a percentage of the previous year’s turnover for serious violations, together with suspension of service, revocation of licences and personal liability for responsible individuals. A 2025 amendment to the Cybersecurity Law, effective from 1 January 2026, raised penalties for network security failures. The generative AI measures add rectification orders, warnings, suspension and takedown. For a consumer product, a suspension order is a more severe outcome than any fine, which is why Chinese compliance work concentrates on avoiding the trigger rather than litigating the penalty.
One system, two document sets
The efficient approach is a single control library with two evidence packs.
Build controls once: data inventory and provenance, model testing and red-teaming, logging and traceability, human oversight, incident response, supplier due diligence, user disclosure and complaint handling. Then map each control to the EU technical documentation structure and to the Chinese filing and self-assessment questions. Most of the underlying work is shared; the differences sit in the applicant, the disclosure audience and the content architecture.
Three decisions cannot be shared and should be made early. Where inference runs, because China’s data rules and the EU’s transfer rules pull in different directions. Who the legal applicant is, because China requires a domestic entity. How content is moderated, because China requires proactive filtering and reporting while Europe requires transparency and rights safeguards. Design these once, correctly, rather than retrofitting them after a filing is refused.
What to do next
- Classify every product against both regimes in one workshop, using the EU annexes and the Chinese filing triggers side by side.
- Confirm whether any use case is prohibited in Europe or has public opinion attributes in China, because those findings change the roadmap.
- Build the shared control library first, then generate the EU technical file and the Chinese filing pack from it.
- Decide where inference and logs sit before signing cloud contracts, since this choice is expensive to reverse.
- Identify the Chinese domestic applicant and its website filing early, because nothing can be filed without it.
- Budget for continuous content moderation in China and post-market monitoring in Europe as permanent functions.
- Track the EU high-risk timetable and Chinese local practice quarterly, because both have moved repeatedly since 2024.
This comparison reflects the position as of late 2026. The EU high-risk timetable has been subject to adjustment proposals and Chinese provincial filing practice varies, so confirm the current requirements with counsel before relying on either regime. For the Chinese filing mechanics in isolation, see Algorithm Filing in China: A Practical Guide.
Frequently Asked Questions↑ Back to article
Does the EU AI Act apply to a company based outside the EU?
Yes. It reaches any provider that places a system on the EU market or puts it into service in the EU, and also providers in third countries where the output is used in the EU. Having no EU establishment does not remove the obligation, although you will need an authorised representative.
Does China have an equivalent to the EU AI Act's risk tiers?
No. China does not classify AI systems into unacceptable, high and limited risk. It regulates by service type and sector instead, using algorithm filing, generative AI filing and security assessment, plus separate rules for medical devices, vehicles, finance and content.
Is a Chinese algorithm filing the same as an EU conformity assessment?
No. A conformity assessment is a provider-led process that produces technical documentation and, for listed systems, a notified body opinion and CE marking. An algorithm filing is an administrative submission to the Cyberspace Administration of China, reviewed by a regulator, with a filing number published.
Which regime is harder for a generative AI company to satisfy?
They are hard in different ways. The EU regime demands evidence, documentation and governance across the product lifecycle. The Chinese regime demands a domestic applicant, pre-launch security assessment and continuous content moderation, with service suspension as a real enforcement outcome.
Can one compliance programme satisfy both the EU and China?
One control framework can, but one document set cannot. The underlying controls on data governance, testing, logging and human oversight overlap heavily, while the evidence packs, the applicant and the disclosure obligations differ. Build a shared control library and two outputs.
Do we need to disclose training data under either regime?
Neither requires publishing your dataset. The EU AI Act requires general-purpose model providers to publish a summary of training content and maintain a copyright policy, while China requires a description of data sources and legality in the filing and security assessment materials.
How large are the penalties under each system?
EU penalties reach tens of millions of euros or a percentage of worldwide turnover for the most serious breaches. Chinese penalties combine statutory fines under the Cybersecurity Law and PIPL with operational sanctions such as rectification orders, service suspension and app removal, which often matter more than the fine.