Skip to content
DCC · DATA COMPLIANCE CHINA China data law, for overseas counsel.
§ AT A GLANCE

The SPC AI Disputes Opinions, on one page.

China's first adjudication rules for AI disputes, drawn as diagrams rather than summarized as text. Every figure carries its article number; the full translation is one click away.

法发〔2026〕10号 · Fa Fa [2026] No. 10

最高人民法院关于依法审理涉人工智能纠纷案件的意见

Opinions of the Supreme People's Court on Trying Cases Involving Artificial Intelligence Disputes in Accordance with Law

Issued
7 Sep 2026
Issuer
Supreme People's Court
Type
Judicial policy document (司法文件)
Size
5 parts · 24 articles
01

Twenty-four articles, five parts

Nine of the 24 articles are tort rules. That is where the Court spent its effort, and where a foreign operator's exposure sits.

  1. Part I · Arts. 1–2 General requirements
  2. Part II · Arts. 3–11 Torts
  3. Part III · Arts. 12–16 Intellectual property
  4. Part IV · Arts. 17–20 Procedure & crime
  5. Part V · Arts. 21–24 Working mechanisms
02

Three principles · Art. 2

以人为本

People-centered

Protect livelihood rights in the digital era; prevent algorithmic discrimination; keep the individual's right to know and to choose in automated decision-making.

支持创新发展

Support innovation

Inclusive and prudent; tolerate failure; equal rules for large and small operators; police AI-enabled monopoly and unfair competition.

筑牢安全底线

Security bottom line

Liability sized to the harm and the risk; open- and closed-source, general- and special-purpose models distinguished; severe on intentional criminal misuse.

03

The default is fault, not strict liability · Art. 3

Start Does a statute expressly impose no-fault or presumed-fault liability?
Yes
Apply that rule. Physical AI products → product liability (Art. 9). Motor vehicles → Civil Code + Road Traffic Safety Law (Art. 11).
No
Fault-based liability, Civil Code Art. 1165(1). The plaintiff must show fault. What counts as fault is listed below.
What goes to fault application scenario degree of autonomy technical & informational transparency potential risk & reach preventive measures taken · technical feasibility user's foresight & control
04

Who answers for what

The Opinions never say "the AI is liable". They name a party each time. Read across a row for one actor's exposure, down a column for one kind of harm.

Party Personality rights Arts. 4–5, 8 GenAI output Art. 7 Product defect Art. 9 AI output & copyright Art. 12 Consumer harm Art. 10 Driving accident Art. 11
AI developer trains or builds the model not named by the article not named by the article not named by the article liable on the article's conditions not named by the article not named by the article
AI service provider runs the service reachable by injunction or notice liable on the article's conditions not named by the article liable on the article's conditions not named by the article not named by the article
AI user / network user prompts, generates, publishes liable on the article's conditions liable on the article's conditions not named by the article liable on the article's conditions not named by the article not named by the article
Producer & seller of a physical AI product not named by the article not named by the article liable on the article's conditions not named by the article not named by the article liable on the article's conditions
Business operator sells goods or services not named by the article not named by the article not named by the article not named by the article liable on the article's conditions not named by the article
Driver of an assisted-driving car not named by the article not named by the article not named by the article not named by the article not named by the article liable on the article's conditions
liable on the article's conditions reachable by injunction (Art. 8) or notice (Art. 7) not named by the article
05

Generative AI gets the safe harbor, by analogy · Art. 7

The Court reasons that a provider cannot pre-screen every output, but can act on notice. So Civil Code Art. 1195 applies to generative AI the way it applies to a platform — with the clock starting at a qualified notice.

Trigger Output infringes reputation, privacy or another personality right Either the model's own hallucination, or a user's malicious prompt.
Qualified notice Rights holder notifies the provider Must include prima facie evidence of infringement and the rights holder's real identity.
Necessary measures Provider acts promptly Stop generating the infringing content; block the generation instruction.
Acted No liability for the output.
Did not act Liable for the harm caused. Where a user induced the output, the user is liable and the provider is joined under Art. 1195.
06

"AI product" means a physical thing · Art. 9

InsidePhysical carrier (实物载体)
  • Intelligent robots
  • Autonomous vehicles, assisted-driving cars
  • Devices with an embedded model

Product Quality Law — producer and seller bear product liability for a defect. Defect is judged by nature and use, self-learning, updates, user control, national and industry standards, and whether scenarios, limitations and foreseeable risks were truthfully explained and clearly warned.

OutsideService without a carrier
  • Chatbots and assistants
  • APIs, model-as-a-service, apps
  • Cloud-hosted generation

Not a "product". The fault rule (Art. 3) and the notice rule (Art. 7) govern instead. This is the holding of the Hangzhou "AI hallucination" judgment, now adopted by the top court.

07

Five personality-rights rules · Arts. 4–5

AI换脸

Face-swap

An unconsented, identifiable virtual likeness that is used or published infringes the rights to name and likeness.

AI拟声

Voice clone

A voice used as training corpus to reproduce timbre, intonation and style infringes the voice right (Civil Code Art. 1023).

操控虚拟形象

Manipulated avatar

Making an identifiable likeness or voice say untrue things that lower social evaluation is defamation.

AI复活逝者

"AI resurrection"

An unauthorized likeness of the deceased lets close relatives sue under Civil Code Art. 994.

网络开盒

Doxxing

Using AI to aggregate a person's public traces to extract or expose private information is a privacy tort (Art. 5).

Art. 8 · Injunction

Where the harm would be irreparable, a court may order the actor to stop — and order the network service provider or generative-AI provider to stop the service, proportionately.

08

Training on public personal information · Art. 6

  1. 1
    Is the personal information lawfully public?Self-disclosed by the individual, or otherwise already lawfully made public.
    No → ordinary PIPL rules
  2. 2
    Is the processing within a reasonable scope for the model?Purpose against model function · type and sensitivity of the information · the context of disclosure and the use a person could reasonably expect.
    No → infringement
  3. 3
    Has the individual expressly refused?
    Yes → infringement
  4. 4
    Is the impact on the individual major?
    Yes → consent required
  5. Generally not an infringement.The first time the top court has said so for model training. Tracks PIPL Art. 27 and Civil Code Art. 1036.
09

Consumers and cars · Arts. 10–11

大数据杀熟 · Art. 10

Algorithmic price discrimination is a tort

Same goods or services, different price or terms by algorithm. Unreasonable when it:

  • substantively limits the right to know, to choose, or to fair dealing;
  • builds terms from the person's preferences, willingness or ability to pay, or browsing history;
  • offends good faith and commercial ethics — unless the reason is legitimate, sufficient and non-discriminatory.
仿冒名人带货 · Art. 10

Fake-celebrity selling is punitive-damages fraud

AI-impersonated endorsements that amount to fraud draw punitive damages under Consumer Protection Law Art. 55.

Autonomous & assisted driving · Art. 11
  • Vehicle defect → producer and seller pay.
  • Defect plus driver fault → both (Civil Code Art. 1172).
  • Overstated automation claims → civil liability to consumers.
  • Courts may order driving event records from whoever controls the data.
10

Intellectual property · Arts. 12–16

12

AI output & copyright

Liability shared by service type, training-data sources, each party's role, measures and profit. A developer pleading non-infringement must produce training sources, process records and the model's operating mode.

13

Open source

Exemption weighed by license type, rights limits, security measures and disclosure. Free modules with disclosed function and risk can carry no liability for downstream misuse.

14

Patents

AI inventions are patentable subject matter; the natural person who made the creative contribution is the inventor; disclosure is judged by the skilled person.

15

Technology contracts

Breach judged against the contract, the nature of AI R&D and whether the developer made reasonable efforts.

16

Data

Compilation works → Copyright Law. Trade secrets → AUCL. Other datasets → AUCL Art. 13. Data poisoning, malicious annotation and adversarial samples attract liability.

11

AI in the courtroom · Arts. 17–20

Art. 19 · Filing AI-generated material
Verifylaws, interpretations and cases cited
Disclosethe AI assistance when filing
Answer for itauthenticity and accuracy are yours
Sanctions ladder for manipulated AI evidence
claim dismissed fine detention criminal liability
Arts. 17–18 · Tools for opaque systems
  • Adverse inference against a party that withholds documents or electronic data without justification.
  • Evidence preservation to fix key technology; investigation on application or the court's own motion.
  • Technical investigation officers, expert assistants, appraisers and people's assessors on technical questions.
  • AI output as evidence is weighed by prompt design, similarity, consistency on repeat, and the model's training and filters.
Art. 20 · Crimes named

fraud · insult · defamation · damaging commercial reputation · infringing citizens' personal information · illegally obtaining computer-system data · obscene materials · defeating an assisted-driving monitor with aftermarket parts

12

Deliberately left blank

The Court says the drafters could not agree, so the Opinions say nothing. Both questions stay with the trial courts for now.

Not decided

Is AI-generated content copyrightable?

Not decided

Is unlicensed training on protected works infringement?

§ SUBSCRIBE

The Monday brief.

One short email every Monday. New briefs on Chinese data-compliance rules from the previous week, with the source law cited.

Opt-in only. Unsubscribe anytime by replying "unsubscribe" to any issue.

SUPPORT DCC

Keep the publication free to read. Suggested support is $19.99, or choose your own amount.

Support →