Independent map & monitor China Est. 2026 AI Governance Observatory

Mapping the governance of artificial intelligence in China

China regulates AI through binding law. Three NPC statutes set the floor, CAC-led measures bind every algorithm, deepfake and generative service, and the 2026 frontier rules reach AI agents, companions and humanoid robots. This site maps that layered system by force and tracks the court cases shaping it.

WAICO founding memberFounding member of the World AI Cooperation Organization, signed in Shanghai on 16 July 2026 by Wang Yi, Member of the Political Bureau of the CPC Central Committee and Minister of Foreign Affairs.
Explore the map Analyze the case tracker
25instruments mapped
12binding, in force
4sector guidance
2draft / pending
7regional / international
01The map

An enforceability spectrum

Twenty-five instruments, stacked by force and listed the way a statute book would list them. The binding stack has three storeys: NPC statute at the base, CAC-led measures on top, frontier rules at the edge. Tap a row for what it does and where it is published.

All Binding Guidance Draft / pending Regional
All types AI-specific Data protection Cyber and crime Digital and telecom
Binding · foundational lawNPC statute · enforceable
Cyber and crimeGeneral catch

Network security obligations, critical information infrastructure protection and data localisation. Every later AI and data rule rests on it.

Source ↗
Digital and telecomGeneral catch

Data classification with tiered security duties and export controls, up to "core state data" under the strictest regime.

Source ↗
Data protectionVia data

The comprehensive privacy statute: informed consent, access and deletion rights, and limits on processing. It governs the personal data behind AI systems.

Source ↗
Binding · AI-specificCAC-led · enforceable
AI-specificAI-native

The first binding national rules anywhere for public-facing generative AI: content compliance, training-data quality and provider accountability for every generative service offered to the Chinese public.

Where neighbouring jurisdictions steer AI with voluntary guidance, this rule binds and carries sanctions.
Source ↗

The first dedicated rules for recommendation algorithms: transparency, user opt-outs, anti-addiction duties and algorithm filing for high-impact services.

Source ↗
AI-specificAI-native

Deepfake and synthetic-media rules: clear labelling, real-name authentication and content review.

Source ↗
Digital and telecomGeneral catch

Bans facial recognition as the sole verification method and cameras in private spaces. Requires privacy impact assessments, and registration once 100,000 identities are stored.

Source pending
AI-specificAI-native

Visible labels plus embedded metadata labels on all AI-generated text, images, audio, video and virtual scenes, backed by mandatory national standard GB 45438-2025.

Source ↗
AI-specificAI-native

A three-tier ethics review for AI research and development, mandatory where AI can mobilise public opinion, decide with high autonomy or integrate deeply with humans. Adds algorithm audits for gig-economy platforms.

Source pending
Binding · frontier (2026)agents · companions · robots
Digital and telecomGeneral catch

The first dedicated rules for AI companions and emotional-interaction services that simulate human personality, thinking patterns and communication styles.

Source pending
AI-specificAI-native

The first national framework for autonomous AI agents: governance principles, nineteen application sectors and safety red lines. A policy instrument, so it directs rather than sanctions.

Source pending
AI-specificAI-native

A national standard system across the humanoid-robot industrial chain, in six domains from brain-like computing to safety and ethics.

Source pending
Sector guidancenon-binding · practically obligatory
AI-specificAI-native

Algorithm transparency, explainability and fairness testing for credit scoring, insurance underwriting and investment AI.

Source pending
AI-specificAI-native

Validation requirements, physician oversight, liability frameworks and patient consent for clinical AI and medical imaging. Accreditation in practice depends on compliance.

Source pending
AI-specificAI-native

Age-appropriate use policies, academic integrity standards, teacher training and student data protection for AI in schools and universities.

Source pending
AI-specificAI-native

The first national framework for measuring and verifying AI systems, so that performance is measurable, comparable and traceable across fourteen sectors.

Source pending
Draft & pendingnot yet in force
AI-specificAI-native

The forthcoming comprehensive statute, expected to consolidate the administrative measures under one risk-tiered law. A scholarly draft circulated in 2024.

Draft: no enacted text yet
Digital and telecomGeneral catch

Rules for AI avatars and virtual humans: biometric consent for likeness replication, personality rights, and a ban on virtual relatives for minors.

Draft for comment
Regional & internationalvoluntary · incl. two China-authored

China's own eleven-point initiative from the 2023 Belt and Road Forum, implemented by the thirteen-point Global AI Governance Action Plan announced at WAIC Shanghai in July 2025, alongside a proposal for a World AI Cooperation Organization headquartered in Shanghai.

Source ↗

The grouping's first standalone leaders-level AI document, adopted at the Rio summit. China is a founding member, so it applies fully.

Source ↗

The first global standard on AI ethics, adopted by all UNESCO member states including China. China has not published a completed UNESCO readiness assessment.

Source ↗

The first summit-level declaration on frontier-AI safety. China is on the official signatory list of 28 countries plus the EU.

Source ↗

The statement on inclusive and sustainable AI. China signed; the US and UK declined.

Source ↗

Annexed to the Osaka leaders' declaration and drawn from the OECD AI Principles. China backs them as a G20 member; it is not an OECD adherent.

Source ↗

Adopted by consensus with the Pact for the Future; China did not dissociate. Shared commitments on digital and AI governance.

Source ↗
02Second reading

Making sense of China's AI governance model

The map ranks instruments by force. This view adds a second question: does an instrument set out principles or prescriptions? In China the crowded corner is binding and prescriptive. Tap any marker.

Prescriptive Principles Nonbinding Binding
Force without detail stays thin. The draft AI Law would plant a consolidating statute here.
Binding and enforceable
Binding, no direct sanction
Sector guidance
Regional / international
Draft / pending

Tap a marker

Twenty-five instruments, two readings

The binding and prescriptive corner is the crowded one: the CAC-led measures put force and detail together, service by service. The sector guidance sits just left of the binding line, nonbinding in form but treated as obligatory. Dashed markers are drafts; the National AI Law would consolidate the whole stack into one statute.

All twenty-five instruments
Compare this model with
03Case compendium

Chinese court rulings that turn on AI

A running record of Chinese decisions where AI is central to the dispute: the first copyright in an AI-generated image, the first AI voice-cloning judgment, the first generative-AI provider held liable for its outputs, and the first criminal conviction of chatbot developers for what their model said.

The pattern: Chinese courts decide these cases mostly under general law, the Copyright Law, the Civil Code's personality rights and the Criminal Law, while the AI-specific measures appear as a standard of care. The Guangzhou Ultraman ruling was the first to apply the Generative AI Measures inside a judgment.
9proceedings tracked
8decided
1on appeal
03The record

Cases and proceedings

Grouped by what the dispute is about. Tap a case for the facts, the legal basis, and the holding. Case numbers from secondary reporting are flagged in the entries.

All Copyright & authorship Voice, likeness & persona Criminal enforcement

What the record shows

The decisions rest mainly on the Copyright Law, the Civil Code's personality-rights articles (990, 1019, 1023) and the Criminal Law, with the internet courts of Beijing, Guangzhou and Hangzhou doing most of the doctrinal work. The CAC's AI-specific measures enter as a duty-of-care yardstick rather than a cause of action. As of mid-2026 no published judgment yet applies the 2025 Labeling Measures.

How this is compiled

Compiled July 2026 from published judgments, court-released typical cases and credible legal reporting, with status noted per case. A record of reported proceedings, not legal advice.

China · AI-IP Index

How China aligns AI and intellectual property

China grants copyright in human-directed AI images and walks AI patents through examination, so its output side is strongly accommodating. The one open question is training data, where a fair-use signal exists in the case law but not yet in statute.

Accommodation 0.17 · Institution 0. Strongly accommodating on protecting AI-assisted output, and institutionally developed. Its one live question is training data.
The framework

What the index measures

The AI-IP Index reads a country on two questions that usually get folded together. Accommodation asks how far the law will go to protect intellectual property that a machine helped produce. Institution asks how much administrative machinery sits behind that law: its guidance, its international submissions, its disclosure duties.

A country can be generous on paper and have built almost nothing, or cautious and highly organised, so the two are scored apart and reported as a pair. Each rests on three questions, scored 0, 0.5 or 1.

Accommodation

How far the law will protect intellectual property that a machine helped make.

  • Q1. Patentability of AI-assisted inventionsWhether an invention developed with AI help, filed with a human named as inventor, can be granted. It decides whether AI-assisted research can be protected at all.
  • Q2. Copyright in AI-generated outputWhether output made with AI attracts copyright when a person directed the work. It sets who, if anyone, owns AI-assisted creative work.
  • Q3. Training and text-and-data-miningWhether copyrighted works can be used to train a model without the owner's permission. This is the main fault line between model builders and rights holders.
Institution

How much administrative machinery sits behind that law.

  • Q4. WIPO Conversation participationWhether the IP office has engaged the WIPO Conversation on IP and Frontier Technologies with a substantive submission. It shows the office is working the question at the international level.
  • Q5. National IP office AI guidanceWhether the office has published dedicated guidance on examining AI-related filings. It tells applicants where they stand before they file.
  • Q6. Disclosure or transparency dutyWhether a filing has to declare that AI was used. It is the office's main lever for provenance.

The scale runs the same way on both axes: 0 is the open or developed end, 1 is the restrictive or absent one, and 0.5 sits between for anything conditional or untested. Read as a pair, the two scores show whether AI-assisted work can be protected here, and whether the office has said so in writing.

Within each axis the three questions carry equal weight. That is a deliberate choice, not an oversight. An AI developer might reasonably value freedom to train above patent inventorship, but weighting is where indices lose their credibility, so the score stays a flat mean and the choice is stated in the open. Each cell also cites a statute section, a case or an office page, so any single score can be challenged without disputing the rest.

Two positions are worth naming. A country can be open on the law but thin on machinery, an open-but-undeveloped stance that bets on ambiguity. Or it can be restrictive on rights yet run an active office, a restrictive-but-developed stance that reflects a settled policy choice rather than a gap. The pair of scores tells those two apart where a single ranking would blur them.

Calibration

Testing the scale against known ground

A rubric is only as good as the results it produces on cases where the answer is already known. These three jurisdictions sit outside this series and are scored with the same six questions, as a check. Their law is the best documented anywhere, so if the scale puts them where the law actually sits, the scale is doing its job. They are reference points, not editions.

United StatesAccommodation 0.33 · Institution 0.17
Q10Q20.5Q30.5Q40Q50Q60.5

Firm judicial settlement at both poles and the weakest statutory position on training. The courts have closed the door on AI inventors and AI authors, while everything on training rides on unresolved fair-use litigation. There is no per se duty to declare AI use in a filing, only a candour duty that bites where the use is material.

Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022); Thaler v. Perlmutter (D.C. Cir. 2025), cert. denied 2 Mar 2026; Bartz v. Anthropic (N.D. Cal. 2025); USPTO revised inventorship guidance, 90 Fed. Reg. 54636 (28 Nov 2025); 89 Fed. Reg. 58128 (17 Jul 2024); 89 Fed. Reg. 25609 (11 Apr 2024).

EuropeAccommodation 0.17 · Institution 0
Q10Q20.5Q30Q40Q50Q60

The most permissive statutory position on training and the strongest transparency duty in force, sitting over an authorship question no court has answered. Read the label carefully: training and transparency run through EU law, but patents run through the EPO under the European Patent Convention, which is not an EU institution and covers 39 states.

Directive (EU) 2019/790, arts. 3 and 4; Regulation (EU) 2024/1689, art. 53(1)(c) and (d); EPO J 8/20 (2021); EPO Guidelines G-II 3.3.1 (2026 edition); Infopaq C-5/08; Painer C-145/10.

United KingdomAccommodation 0.33 · Institution 0.67
Q10Q20.5Q30.5Q40.5Q50.5Q61

Liberal on patents since the Supreme Court abandoned the Aerotel test in February 2026, conservative on copyright, and currently thin on AI-specific machinery after the dedicated AI examination guidance was withdrawn. Section 9(3) nominally protects a work with no human author for 50 years, the widest AI-output right on paper, but no court has ever applied it to generative AI.

CDPA 1988 ss. 9(3), 12(7), 29A, 178; Thaler v Comptroller-General [2023] UKSC 49; Comptroller-General v Emotional Perception AI [2026] UKSC 3; UKIPO practice notice, 14 Jul 2026; Data (Use and Access) Act 2025 ss. 135 to 137.

The check earned its keep twice. The untested rule held: section 9(3) of the UK Copyright Act reads like the widest AI-output right anywhere, yet because no court has applied it to generative AI it scores as untested rather than open, which is the correct answer and the one a headline reading would miss. The check also exposed a limit worth stating. The Institution axis measures AI-specific machinery, not general office capacity, so a long-established office can score poorly when it has withdrawn its AI guidance and imposes no disclosure duty. Read that axis as what an office has built for AI, nothing wider.

Scorecard

The evidence, question by question

Each row below carries the test we applied and the finding behind the score, with a source. The rationale, not the number, is the point.

Direction of travelLoosening. Courts are trending toward allowing training on copyrighted works, and iQIYI v. MiniMax alongside the Supreme People's Court 2026 to 2030 IP plan will test whether that fair-use signal hardens into settled law. This is recorded apart from the scores: a pending change is signal, not yet the rule.

Scores verified 16 July 2026, against AIP Index codebook v1.0. A score is only true as of its date.

0 · open or developed0.5 · conditional or emerging1 · restrictive or absent

Accommodation: what the law permits

Mean 0.17
0
Q1. Patentability of AI-assisted inventions

What we look forA granted patent or a clearly workable examination route, not merely a silent statute.

The findingCNIPA's December 2024 guidelines walk AI-related inventions through examination with a human named as inventor, and the office grants on that basis. A workable, applied rule.

CNIPA AI patent guidelines
0
Q2. Copyright in AI-generated output

What we look forProtection applied where human choices shape the work, and refusal of purely machine output.

The findingCourts from Beijing to Changshu and Wuhan protect human-directed AI images where the prompter's choices show original expression. The rule is applied, not theoretical.

Li v. Liu and later rulings
0.5
Q3. Training and text-and-data-mining

What we look forA statutory or judicial mining lane. Silence scores against the miner, since the reserved-rights default governs.

The findingThere is no statutory mining exception, but the Hangzhou Ultraman appeal said training on copyrighted works can be fair absent intent to plagiarise or market harm. An emerging judicial approach, not settled law, and iQIYI v. MiniMax will test it.

Hangzhou Ultraman appeal

Institution: how developed the machinery is

Mean 0
0
Q4. WIPO Conversation participation

What we look forA traceable, substantive submission on the record, not bare membership.

The findingChina and CNIPA are documented active participants in the WIPO Conversation on IP and Frontier Technologies.

WIPO Conversation records
0
Q5. National IP office AI guidance

What we look forA standalone AI examination text, not AI handled quietly under general practice.

The findingCNIPA has published dedicated AI patent examination guidelines, a rare piece of formal office-level guidance on AI.

CNIPA guidelines, Dec 2024
0
Q6. Disclosure or transparency duty

What we look forA mandatory declaration inside the filing, not a proposal or an informal request.

The findingCNIPA's AI patent guidelines require the training datasets and training process to be disclosed in the specification, a training-data disclosure inside the filing, framed for sufficiency rather than provenance.

CNIPA AI patent guidelines

Conclusions

Can AI be named as inventor?
No, a machine cannot be named
Statutory opt-out from training?
None in statute
Volume of disputes
A developing body of cases

Instruments where AI meets IP

Copyright

Copyright Law (2020 amendment)

The open door courts walked through
NPC · In force Jun 2021
No AI text, but the open-ended definition of works as original intellectual achievements is the hook courts use to protect AI-generated output, with authorship confined to humans and legal entities. Li v. Liu's four-element test applies it.
Generative AI

GenAI Interim Measures, Article 7

IP compliance at the training stage
CAC and six co-regulators · In force Aug 2023
Training must use data and foundation models from lawful sources and must not infringe others' intellectual property rights, with a parallel duty to respect IP in providing and using the service.
Patents

CNIPA AI Patent Guidelines (Trial)

Facilitating AI patents, humans only
CNIPA · Adopted 31 Dec 2024
Four categories of AI-related inventions with rules on eligibility, disclosure, inventiveness and ethics, designed to help applications through. The inventor must be a natural person; naming an AI is barred, codifying the DABUS rejection.
GuidelinesAnalysis ↗
Synthetic media

Deep Synthesis Provisions, Article 6

IP under the general clause
CAC, MIIT, MPS · In force Jan 2023
Bars using deep synthesis to produce or spread information that infringes others' lawful rights and interests. The IP coverage is a general-clause prohibition rather than a dedicated IP article.
BindingSource ↗
Judicial policy

SPC IP Protection Plan 2026-2030

The codification commitment
Supreme People's Court · 20 Apr 2026
Commits the courts to defining the legal attributes of AI-generated content, rules for training corpora, and liability allocation among developers, operators and users, with adjudication opinions on AI disputes in drafting.
Judicial policyAnalysis ↗
Statute

AI Law (scholars' suggestion draft)

Inputs and outputs as IP, if enacted
Academic draft · Mar 2024, no official bill yet
The scholars' draft would permit copyrighted material in training for most cases and extend IP protection to AI-assisted content. It is not an official government bill; the AI Law remains a legislative-plan item.
Unofficial draftTranslation ↗

The case record

The founding four, Li v. Liu, Dreamwriter and the two Ultraman rulings, live on the Cases tab under copyright and authorship. These are the AI-IP decisions beyond them.

Holding

An image made with Midjourney and refined in Photoshop was protected, and defendants who copied it into advertisements and a three-dimensional installation infringed. The second Chinese court to protect AI-generated imagery.

Holding

The first Chinese refusal of copyright in AI output: butterfly-chair images lacked sufficient human input, and prompts themselves are not copyrightable. Protection tracks the human contribution, not the tool.

Holding

Copyright can subsist in AI images, but the claimant must evidence the creative effort: prompt records and the selection and modification process. The doctrine matures from whether to how to prove it.

Holding

An AI one-click video app that cut and spliced episodes of the drama Joy of Life and induced users to generate and spread infringing clips was held liable for RMB 800,000. A case about an AI tool infringing a drama's rights, not about protecting AI output.

What it is

The first suit by a Chinese video platform against an AI developer over unauthorized use of films and shows for model training, plus infringing outputs. Still pending as of mid-2026; the decision will test the Hangzhou training-can-be-fair dicta against a live claim.

What it is

CNIPA rejected the AI-as-inventor application and upheld the rejection on re-examination, a decision featured among its top re-examination cases. The inventor's administrative suit sits at the Beijing IP Court with no reported judgment as of mid-2026.

Method. Six questions, three per axis, scored 0, 0.5 or 1 against AIP Index codebook v1.0. The three questions in each axis are weighted equally by choice, and the two axes are reported as a pair, never blended into one number. Every score cites a statute, case or office page, carries a verified date, and records its direction of travel apart from the level. Positions are editorial judgements read from primary law and office practice, not official scores. WIPO's public catalogue of IP-office AI initiatives does not score jurisdictions on this basis, so the readings here come straight from the source law. Not legal advice.

04About

Why this exists

China regulates AI the other way round from most of the region. Its neighbours steer AI with voluntary guidance while binding duties come only from general law; in China the AI-specific rules are the binding ones, issued measure by measure as each technology arrived: algorithms in 2022, deepfakes and generative AI in 2023, biometrics and labeling in 2025, agents and companions in 2026.

The site does two things: it maps that layered system as an enforceability spectrum, from NPC statute down to sector guidance, and it tracks the court rulings where AI is central, a docket that includes several world firsts. The domestic instrument set follows the China AI Law Explorer's inventory of official sources, extended with China's verified international commitments; case entries are drawn from published judgments and court-released typical cases.

The structure is easy to hold in your head: three NPC statutes at the base, CAC-led measures on top, frontier rules at the edge, and a draft National AI Law that would pull all of it into one statute.

Behind this country report

Muhammad Deckri AlgamarIP & digital law · LL.M. (IE-WIPO), AIGP
Prof. Abu Bakar MunirCyber law & data protection · Advisor on Malaysia's PDPA 2010, DPEX Chairman, APPDI Co-founder

Contribute

Know of a Chinese ruling or proceeding where AI is central? Send a source link; it will be reviewed before publishing.

Submit a case
05Outlook

What comes next

China is consolidating. The measure-by-measure approach built a dense stack of binding rules, and the draft National AI Law is expected to pull them under one NPC statute with risk-tiered classification. The 2026 frontier rules show the pattern continuing at the edge: agents, companions, humanoid robots and virtual humans each get their own instrument as the technology lands.

The internet courts in Beijing, Guangzhou and Hangzhou keep producing first rulings, and judges have started reading the CAC measures as a standard of care rather than a regulator's text alone. This site keeps the map and the record current, so the consolidation is easy to follow.

An independent research and educational project. Information is provided as-is, may be incomplete, and does not constitute legal advice. Case entries record reported proceedings on a stated date, drawn from public sources, not a legal finding of guilt or liability.