Part of the AI Regulation News hub.
China's Supreme People's Court has issued 24 articles directing courts at every level to decide AI disputes on a fault liability default, and to order a developer that denies copyright infringement to produce its training data provenance and model operating mode
The Opinions are a non-binding judicial policy document addressed to the courts, and by their terms they write no new statutory duty onto any private party. The headline fact is the liability default. The provision that will change behaviour is a production direction aimed at developers. Then there are the two questions the 24 articles decline to answer at all.
Bottom line: Issued and signed 7 September 2026 by the Supreme People's Court and numbered 法发〔2026〕10号. It is a judicial policy document addressed to the courts. It is not a statute, and it does not write new duties directly onto private parties. The text we read carries no commencement clause.
Who this affects: Litigation counsel appearing before PRC courts, intellectual property and personal information compliance leads at model developers and generative AI service providers operating in China, and product counsel at assisted and autonomous driving manufacturers.
Issue date: Signed 7 September 2026 in the signature block at the foot of the operative text. The Supreme People's Court page carries a publication timestamp of 7 September 2026 at 19:22. No deadline runs from the document.
What changed: China's highest court has told every level of court how to approach AI disputes: which liability standard applies by default, how to treat already public personal information used in training, what a court is to order from a developer that denies copyright infringement, and which AI things count as products. It does that in a judicial policy document that directs the courts and creates no new statutory duty on private parties.
Analysis: Two of the loudest questions in AI copyright are absent from all 24 articles. The court's own introduction says issues without consensus were left blank. It does not say which ones, so we read the operative text and report the gaps as gaps.
Primary sources: Opinions of the Supreme People's Court on Adjudicating AI-Related Dispute Cases in Accordance with Law, full text
- Instrument (EN)
- Opinions of the Supreme People's Court on Adjudicating AI-Related Dispute Cases in Accordance with Law
- Authority
- Supreme People's Court of the People's Republic of China
- Jurisdiction
- China, national. Addressed to people's courts at every level
- Status
- Issued and published 7 September 2026. Five parts, 24 articles
- Bindingness
- A judicial policy document styled Opinions and numbered in the 法发 series. It directs the courts. It does not by its terms impose a new statutory duty on any private party, and we do not state whether it may be cited as the legal basis of a judgment
- Issue date / next deadline
- Signed 7 September 2026. No deadline and no transition period appear in the text
- Legal basis
- Named in the preamble: the Civil Code, the Cybersecurity Law, the Data Security Law, the Copyright Law, the Anti-Unfair Competition Law, the Consumer Rights and Interests Protection Law, the Personal Information Protection Law and the Civil Procedure Law
- Document
- 法发〔2026〕10号
- Primary source
- https://www.court.gov.cn/zixun/xiangqing/511101.html
The default is fault liability, and the court lists what to weigh
Article (3) sets the starting point for everything that follows it. Where the law does not expressly provide for no-fault liability or for presumed fault, the Opinions direct courts to decide whether a defendant bears tort liability under the fault principle in article 1165(1) of the Civil Code.
The Chinese text reads: 法律没有明确规定适用无过错责任或者过错推定责任的，应当依照民法典第一千一百六十五条第一款规定的过错责任原则认定行为人是否承担侵权责任。
Our translation, offered as a translation and not as the operative text: where the law does not expressly provide for the application of no-fault liability or presumed-fault liability, whether the actor bears tort liability shall be determined under the fault liability principle provided in article 1165, first paragraph, of the Civil Code.
The same article then sets out what a court weighs in deciding whether there was fault and how serious it was. Our translation of that list: the specific application scenario of the AI, the degree of autonomy, technical and information transparency, the potential risks and the scope of their impact, the measures that developers, providers and other relevant parties took to prevent and reduce AI infringement together with what was technically possible, and the user's ability to foresee and control the harm that its use of the technology might cause.
Set that against the alternative and the significance is easier to see. A no-fault or presumed-fault default would have put developers on the defensive in every case that reached a court. Article (3) goes the other way, then supplies a list a defendant can build a record against before any dispute exists.
What a developer is to put on the table when it denies infringement
Article (12) sits in the intellectual property part of the Opinions and deals with AI-generated content that infringes another party's copyright. It tells courts to weigh the type of AI service, industry characteristics, the source of training data, the degree of participation of each party, the measures taken and the profits obtained, and then allocate responsibility among developer, provider and user.
The sentence that matters most to a developer is narrower: 人工智能开发者提出不侵权抗辩的，应当责令其提供训练数据来源、训练过程记录、模型运行模式以及科学理论依据等予以佐证。
Our translation: where an AI developer raises a non-infringement defence, it shall be ordered to provide the source of the training data, records of the training process, the model's operating mode and the scientific theoretical basis, in support of that defence.
Two limits belong with that sentence. It sits in the copyright part and is framed around generated content infringing copyright, so it is not a general disclosure rule for every AI dispute. And the sentence immediately after it runs the other way: where a rights holder alleges that an AI provider used algorithmic technology to infringe copyright, the rights holder is the one to provide the relevant evidence.
The practical consequence for a developer is about records, not pleadings. Training data provenance, training process records and model operating mode either exist in a retrievable form or they do not, and that is settled long before anyone files.
Training on information that is already public
Article (6) is the provision most likely to be quoted out of context, so here is the operative sentence: 为人工智能模型训练，在合理范围内处理个人自行公开的或者其他已经合法公开的个人信息，且个人未明确拒绝的，一般不认定为侵害个人信息权益的行为。
Our translation: where, for the purpose of AI model training, personal information that the individual has made public themselves or that has otherwise already been lawfully made public is processed within a reasonable scope, and the individual has not expressly refused, this is generally not to be found to be conduct infringing personal information rights and interests.
Then the carve-out, which is the half that gets dropped in summaries. Our translation of the next sentence: where there is a material effect on the individual's rights and interests, consent shall be obtained in accordance with the law. The text does not define material effect.
Reasonable scope is not left at large either. The article lists the purpose of the processing measured against the necessity and appropriateness of the model's function, the type and sensitivity of the personal information and its potential impact on the individual, and the context in which the individual made the information public together with the scope of use that could reasonably be expected from that context.
Two doctrines moved onto AI: notice and necessary measures, and the meaning of a product
Article (7) deals with generative AI service providers. Where automatically generated content infringes reputation, privacy or other personality rights and interests, and the provider does not promptly take necessary measures such as stopping the generation of the infringing content after being notified by the rights holder, it bears tort liability for the resulting harm. The notice is to include preliminary evidence of infringement and the rights holder's true identity information.
The second half of that article addresses the user who goes looking for trouble. Where a network user maliciously induces a generative AI system to produce infringing content by inputting infringing prompts and causes harm, the user bears liability. Where the provider then fails to take necessary measures after notice, including stopping the generation and blocking the relevant generation instructions, and the rights holder brings a claim under article 1195 of the Civil Code against the user and the provider, the Opinions direct courts to support it.
Article 1195 is the Civil Code's notice and necessary measures provision, written for network service providers. Routing a claim against a generative AI service provider through it is the move worth noting, and describing that as extending a safe harbour by analogy is our characterisation rather than language the Opinions use.
Article (9) does something quieter and possibly more consequential. It directs courts to apply the definition of product in the Product Quality Law and to identify AI products 以实物为载体, that is, carried on a physical object. Where such a product is defective and causes harm, producers and sellers bear product liability. In assessing unreasonable danger to person or property a court is to weigh the product's nature and use, its autonomous learning capability, its upgrade and update history, the degree of user control over the system and conformity with national and industry standards, with particular attention to whether producer and seller gave truthful explanation and clear warning about application scenarios, inherent limitations and foreseeable risks.
What that phrase implies for a service with no physical carrier is left to be worked out. We read it as steering purely software-delivered AI services away from the product liability route, and we flag that as our reading, because article (9) states the inclusion and does not state the exclusion.
If you file in a PRC court, article (19) is the one to circulate
Article (19) is the one place where the Opinions speak to a filing party rather than telling a court what to do, and they do it inside a judicial policy document rather than a statute. Our translation of the final sentence of article (19): where litigation participants submit pleadings, case search reports or other materials that were generated using artificial intelligence, they shall, before submitting them to the court, carefully verify the authenticity and accuracy of the laws, judicial interpretations and cases they contain, shall explain the use of AI assistance when submitting them, and shall bear responsibility in accordance with the law for the authenticity and accuracy of that content.
Note what that sets out. Verification, an affirmative disclosure to the court that AI assistance was used, and responsibility for the content, all three attaching to the person who files.
The rest of article (19) is harder edged. Where a party obtains false evidence by human intervention or misdirection, including deleting or altering generated-content labels, feeding specific instructions, presenting results selectively or using adversarial interference, and fabricates the basic facts of a civil case to bring a false action, courts are directed to reject the claim and, depending on severity, to impose fines or detention, with criminal liability where the conduct constitutes a crime. Forging evidence by AI to obstruct a court's handling of a case is routed to article 114 of the Civil Procedure Law.
The two questions the 24 articles do not answer
The introduction published by the court's news bureau alongside the text says the Opinions left blanks deliberately: 对于目前难以形成共识的问题，《意见》做了留白处理，待进一步积累经验，条件成熟时再通过适当方式作出明确。 Our translation: for issues on which consensus is currently difficult to reach, the Opinions adopt a blank-space approach, to be clarified in an appropriate manner once further experience has accumulated and conditions are ripe.
That statement names no issue. So we read the 24 articles and report the absences as absences. There is no provision on whether AI-generated content is itself a copyrightable work, and no provision on whether using copyrighted works without licence to train a model is fair use. The term 合理使用 does not appear in the operative text.
The contrast inside the same part of the document is what makes the omission legible. Article (14) does answer the equivalent question on the patent side: an AI-related invention that uses technical means following natural laws, solves a technical problem and achieves a technical effect consistent with natural laws is to be recognised as patentable subject matter, subject to exclusions, and a natural person who has made a creative contribution to the substantive features of an invention completed using AI is to be recognised as the inventor.
Pairing the acknowledged blanks with these two specific gaps is our inference from reading both documents. The court did not say the blanks are those two things.
What we did not verify
What we opened: the operative text of 法发〔2026〕10号 on the Supreme People's Court site, read end to end in Chinese, covering all five parts and all 24 articles, including the signature block reading 最高人民法院 above the date 2026年9月7日. We also read the introduction published by the court's news bureau on the same page, which is where the five-part, 24-article structure, the blank-space statement and the sentence 目前我国尚未出台专门的人工智能法。 come from. Our translation of that last one, offered as a translation and not as the operative text: at present our country has not yet promulgated a dedicated artificial intelligence law.
What we did not open: the question-and-answer piece the page links at /zixun/xiangqing/511111.html, two follow-on newspaper pieces we located elsewhere on the court's site at /zixun/xiangqing/511171.html and /zixun/xiangqing/511271.html, the provisions of the Civil Code, Product Quality Law, Personal Information Protection Law and Civil Procedure Law that the Opinions cite, and any decided case applying them. We describe article 1195 and article 1165 only as this document invokes them.
What we refuse to claim: we do not call this legislation, and we do not call it a judicial interpretation, because the document styles itself Opinions. We do not state whether it may be cited as the legal basis of a judgment, because that turns on rules we did not open. We give no commencement date, because the text we read carries no commencement clause. We do not say the court identified which issues it left blank, because its introduction does not name them. We do not say that AI services without a physical carrier are excluded from product liability, only that article (9) directs courts to identify AI products carried on a physical object. The court's own introduction describes the document as the first judicial adjudication rules document on AI issued by the national highest adjudication organ, and that is its characterisation, reported as such.
Every English rendering of the Chinese text in this piece is our translation and is labelled as one. Quoted Chinese is reproduced as published, including its punctuation.
Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.
If you develop or serve models into China, the action item is evidentiary rather than legal. Article (3) gives you a fault standard and a published list of what a court weighs, and article (12) tells a court to order a developer raising a non-infringement defence to produce training data provenance, training process records, model operating mode and a scientific basis. Both of those reward a firm that can retrieve that record on demand and punish one that cannot. Build the record now, because the defence that depends on it is the one you will want later.
Source File
https://www.court.gov.cn/zixun/xiangqing/511101.html
Open the operative text on the Supreme People's Court site and confirm four things: the fault liability sentence in article (3), the production direction in the second sentence of article (12), the physical carrier wording in article (9), and the signature block dated 2026年9月7日 sitting below all 24 articles and below the document number 法发〔2026〕10号.
法律没有明确规定适用无过错责任或者过错推定责任的，应当依照民法典第一千一百六十五条第一款规定的过错责任原则认定行为人是否承担侵权责任。 · 最高人民法院关于依法审理涉人工智能纠纷案件的意见, article (3), 7 September 2026
FAQ
Is this China's AI law?
No. China has not enacted a dedicated AI statute, and the introduction published with this document says so. This is a judicial policy document styled Opinions, numbered 法发〔2026〕10号, which tells the courts how to apply existing statutes to AI disputes. The statutes it names as its basis include the Civil Code, the Copyright Law and the Personal Information Protection Law.
Does it reverse the burden of proof against AI developers?
Not generally. Article (3) keeps fault liability as the default. The production direction in article (12) is narrower: it applies where an AI developer raises a non-infringement defence in a copyright matter, and the sentence after it puts the evidential burden on a rights holder who alleges that a provider infringed copyright through algorithmic technology.
Does it settle whether AI-generated output is copyrightable?
No. We read all 24 articles and there is no provision on the copyrightability of AI output, and none on whether unlicensed use of copyrighted works to train a model is fair use. The court's introduction says issues lacking consensus were left blank without naming them; pairing that statement with these two gaps is our reading.
What does it say about assisted and autonomous driving?
Article (11) routes crash liability through the Civil Code and the Road Traffic Safety Law, supports product liability claims against producers and sellers where a vehicle defect caused the harm, and supports joint claims against driver and manufacturer where a defect combined with driver fault produced the same harm. It also addresses false or misleading marketing of automation level and performance, and it lets a court require producers, sellers or operators holding the data to provide genuine and complete event recorder data within the scope necessary to establish the cause.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.