China's Supreme People's Court Opinion on Adjudicating AI-Related Disputes: Key Points

On September 7, 2026, the Supreme People's Court issued its Opinion on Adjudicating Disputes Involving Artificial Intelligence in Accordance with Law (Fa Fa [2026] No. 10), together with a press conference the same day. According to the Court's press office, the conference was chaired by spokesperson Ji Zhongbiao; Vice President Tao Kaiyuan announced the Opinion; and Zhou Jiahai (Director of the Research Office), Li Jian (Presiding Judge of the Third Civil Division), and Si Yanli (Deputy Director of the Research Office) attended and took questions.

The Opinion has five parts and 24 articles. In its release materials the Court describes it as "the first document of judicial adjudication rules on artificial intelligence issued by the highest adjudicative authority of the State"; on questions where consensus is not yet possible, the Opinion "leaves matters open, to be clarified through an appropriate instrument once experience has accumulated and conditions are ripe."

China has not yet enacted a dedicated artificial intelligence statute. The Opinion's approach is to give guidance on applying existing law — the Civil Code, the Cybersecurity Law, the Data Security Law, the Copyright Law, the Anti-Unfair Competition Law, the Consumer Rights Protection Law, the Personal Information Protection Law, and the Civil Procedure Law — to the dispute types already reaching the courts. What follows tracks the Opinion's five parts. The full text appears in our same-day republication.

I. General Requirements: Three Basic Principles (Articles 1–2)

Part One establishes three principles: people-centered; supportive of innovation and development; and firm on the security baseline.

The section on the security baseline sets out the basic approach to allocating liability — "distinguishing the differences in technical principles, risk externalities, and control capability among different types of large models, including general-purpose versus special-purpose and open-source versus closed-source, and reasonably allocating the legal liability of developers, providers, users, and other actors." The specific rules that follow can mostly be traced to provisions in Parts Two and Three.

II. Tort Liability: From the Basis of Liability to Specific Scenarios (Articles 3–11)

Fault liability as the basis

Article (3) provides that "where the law does not expressly provide for no-fault liability or presumed fault, whether an actor bears tort liability shall be determined under the fault liability principle in Article 1165, paragraph 1, of the Civil Code." In its answers to press questions, the Court stated the rationale plainly: "to avoid imposing excessive liability while artificial intelligence technology and industry remain at an early stage of development, which would dampen the enthusiasm for innovation."

The provision lists three groups of factors relevant to fault: the specific application scenario, degree of autonomy, technical and informational transparency, and the potential risk and scope of impact; the measures developers and providers took to prevent and reduce infringement and what was technically possible; and users' ability to foresee and control the harm the conduct might cause.

AI face-swapping and voice cloning; digital resurrection of the deceased

Article (4) proceeds in four layers. Processing a natural person's name, likeness, or the like with AI without consent to generate and use or publish a virtual digital representation identifiable as that person infringes personality rights including the rights to name and likeness. Using a natural person's voice as training material without consent to imitate that person's timbre, intonation, and pronunciation style and generate an identifiable synthetic voice infringes voice rights. Manipulating the generated representation or voice to engage in improper conduct or make false statements that lower social evaluation infringes reputation rights. Unauthorized creation or use of a deceased person's virtual digital representation that infringes their name, likeness, or reputation permits close relatives to seek civil liability under Article 994 of the Civil Code.

The training-material layer targets the act of "using a natural person's voice as training material" itself, and does not depend on whether the generated output is published.

Doxxing and human flesh search

Article (5) treats the following as infringement of privacy: using AI, "for the purpose of prying into privacy," to track and analyze a specific natural person's telephone number, online accounts, social media, and other public information so as to obtain private information; or disclosing or publishing private information so obtained; or using such information to disturb the tranquility of private life.

The focus is aggregating private information out of public information. That each individual item is already public does not affect the characterization of the aggregation.

Personal information in training data

Article (6) draws a boundary: processing, within a reasonable scope and for model training, personal information that an individual has made public or that has otherwise been lawfully made public, where the individual has not expressly refused, is generally not an infringement of personal information rights; where there is a material effect on individual rights, consent must be obtained as the law requires.

"Reasonable scope" turns on the necessity and appropriateness of the processing purpose relative to the model's function; the type and sensitivity of the personal information involved and the potential effect on individual rights; and the context in which the individual made the information public and the scope of use reasonably to be expected. All three point to the same question: when the individual made the information public, could they reasonably have foreseen its use for model training?

Liability of generative AI service providers

Article (7) establishes a notice-and-necessary-measures mechanism: where generated content infringes personality rights such as reputation or privacy and the provider, having been notified by the rights holder, fails to take necessary measures such as ceasing generation of the infringing content in a timely manner, the provider bears tort liability for the resulting harm. The notice must include preliminary evidence of infringement and the rights holder's true identity.

Where a network user maliciously induces the generation of infringing content by entering infringing prompts and causes harm, that user bears tort liability; and where the provider, having been notified, fails to take necessary measures such as ceasing generation or blocking the relevant instructions, the rights holder may seek civil liability from the user and the provider under Article 1195 of the Civil Code.

The Court's answers to press questions explain this provision more clearly than the text itself. The Court noted that in practice "there is disagreement over whether generative AI service providers may invoke the safe harbor rule by analogy." Its reasoning is that generated output depends on the training data, model parameters, and the user's prompt, and that a provider "cannot foresee in advance everything network users will enter," so it "cannot be required to predict, review, and intercept every generated output for infringement" — but once notified, "it should, and is able to, take necessary measures." On that basis the Court found a "legitimate and reasonable basis for applying the safe harbor rule by analogy."

Personality rights injunctions

Article (8) brings the Civil Code's personality rights injunction into this area: where an applicant has evidence that an actor is engaging in or about to engage in unlawful conduct infringing their personality rights, and a failure to stop it promptly would cause irreparable harm to their lawful rights, the applicant may seek an order requiring the actor to cease the conduct, or requiring a network service provider or generative AI service provider to cease providing the relevant service. In issuing such an order, a court should consider the type of personality right infringed, the manner of the unlawful conduct, and the scope and degree of potential harm, and "shall not exceed what is necessary."

AI product liability

Article (9) requires that an AI product be identified under the definition of "product" in the Product Quality Law, with the text expressly directed at "AI products with a physical carrier." The Court's answers go further: limiting AI products "to products with a physical carrier — intelligent robots and autonomous vehicles, for example — and excluding AI services without a physical carrier," so as to "avoid an overbroad expansion of product liability." Purely software-form AI services therefore do not proceed under product liability and must return to the fault liability of Article (3).

In determining whether an unreasonable danger exists, courts consider the product's nature and use, autonomous learning capability, upgrade and update history, the degree of user control over the system, and conformity with applicable national and industry standards, and "shall focus on whether the producer and seller gave truthful explanation and clear warning" regarding the product's application scenarios, inherent limitations, and foreseeable risks.

Algorithmic price discrimination; celebrity impersonation in livestream sales

Article (10) addresses algorithmic differential pricing, with three criteria: whether there is a substantive restriction on or harm to consumers' rights to information, free choice, and fair dealing; whether transaction terms are individualized based on the consumer's preferences, willingness to pay, ability to pay, browsing history, and similar information; and whether the conduct violates good faith and commercial ethics. The same article provides that where a business uses AI to impersonate a celebrity in livestream selling and the conduct constitutes fraud, a court will support a consumer's claim for punitive damages under Article 55 of the Consumer Rights Protection Law.

Autonomous and assisted driving

Article (11) distinguishes several situations: where a product defect in the vehicle causes an accident, the producer or seller may be liable under Book Seven, Chapter Four of the Civil Code; where a vehicle defect combines with driver fault in an assisted-driving vehicle to cause the same harm, both the driver and the producer or seller may be sued under Article 1172 of the Civil Code; and where false or misleading claims are made about the level of automation, degree of intelligence, performance, or use, consumers may seek civil liability from the producer or seller.

The article's final sentence concerns evidence: to determine the cause of an accident, a court "may require the vehicle producer, seller, operator, or other data controller to provide, within the necessary scope, truthful and complete autonomous and assisted driving event records and other data needed to establish the facts of the case."

III. Intellectual Property (Articles 12–16)

Allocating the burden of proof where generated content infringes copyright

Article (12) identifies as relevant factors the type of AI service, industry characteristics, the source of the training data, each party's degree of participation, the measures taken, and the profits obtained. The Court's answers tie these to one theme — "liability should match control capability and duty of care" — and frame the training data question as "which party did the 'feeding.'"

The allocation of the burden of proof is the most practical content in this part. The provision states that "where an AI developer asserts non-infringement, 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 basis in support"; and that a rights holder asserting that a provider infringed its copyright through algorithmic technology shall provide supporting evidence. The Court's answers set out both thresholds in more detail: the rights holder should make a preliminary showing that "the infringing content was generated by that AI and is substantially similar to its work"; while the developer, because "AI technology is complex and opaque and the relevant evidence is not obtainable by others," should bear the burden on the training data source, training process, and operating mode, and where necessary provide the scientific basis.

A separate rule applies to users: one who knows or should know of a prior work and uses AI to generate a work substantially similar to it, without a reasonable defense, may be held liable at the rights holder's request.

The Two Questions Left Open

The Opinion does not answer two questions, which the Court's answers name expressly: the copyrightability of AI-generated content, and the characterization of using another's work to train a large model without permission. The Court explained that during drafting there were "sharp differences of view" on both and that "understanding requires further development," so "the Opinion does not address these two questions for the time being."

These are precisely the two most contested points in China and abroad. Article (12) addresses how the parties allocate liability where generated content infringes another's copyright — which is distinct both from whether the generated content can itself be a work and from whether training on works is lawful. That line is worth drawing clearly when citing the Opinion.

Open-source liability exemption

Article (13) gives open-source contributors a relatively clear path: where a party provides, free and open-source, part of the code modules needed for AI software development, and publicly explains their function and security risks, and another party's use of those modules results in infringement, a court "may find that the open-source software developer or provider does not bear tort liability."

The exemption is not automatic. The provision conditions it on four considerations: the type of open-source license, the specific content of the rights limitations, the security and compliance measures taken, and the degree of information disclosure. The requirement to "publicly explain function and security risks" turns on the open-source party's own disclosure.

Patentability and inventorship

Article (14) resolves three questions in three sentences. Patentability: an AI-related invention that employs technical means following natural laws, solves a technical problem, and achieves a technical effect consistent with natural laws is patentable subject matter — except where it violates law or social morality, harms the public interest, or where no natural person made a substantive contribution. Inventorship: where a natural person uses AI to complete an invention and made a creative contribution to its substantive features, that natural person is recognized as the inventor. Sufficiency of disclosure: where the specification describes the technical solution to a degree enabling a person of ordinary skill in the art to carry out the invention, the sufficiency requirement is satisfied.

Tiered protection of data rights

Article (16) distinguishes several layers: data lawfully obtained and in which rights are held is protected; data constituting a compilation work or otherwise meeting the requirements of a work is protected under the Copyright Law; data constituting a trade secret is protected under the Anti-Unfair Competition Law; and where data does not constitute a trade secret but the challenged conduct violates Article 13 of the Anti-Unfair Competition Law, liability follows. The article further provides for liability where data and algorithms are used to reach monopoly agreements or abuse market dominance, and where fabricated interference data, malicious data labeling, adversarial example attacks, or similar means harm the operational security of AI systems.

On unfair competition, the Court's answers gave concrete examples: using AI-generated text, images, video, or virtual characters to fabricate traffic and favorable reviews, and "using AI face-swapping to fabricate explainer videos for false advertising" — treated as passing off and false advertising.

Article (15) separately sets out principles for breach of contract liability in technology development, transfer, licensing, consulting, and service contracts, requiring consideration of the characteristics of AI research and development and whether the developing party made reasonable efforts.

IV. Procedural Rules and Criminal Liability (Articles 17–20)

Fact-finding and review of evidence

Article (17) lists the available tools: litigation guidance and clarification, party applications for court-ordered evidence collection, court collection on its own authority, evidence preservation, and the consequence of a document production order — where a party controlling documentary or electronic evidence refuses without justification to produce it, and the opposing party asserts that its contents are unfavorable to the controlling party, the court may find that assertion established. Where technical principles, operating mechanisms, or other specialized questions are involved, the court may draw on people's assessors, appraisers, expert assistants, and technical investigators.

Article (18) sets differentiated review priorities by evidence type: for big-data analysis reports, the source of the underlying data, the cleaning rules, and the scientific soundness of the analytical method; for blockchain-preserved evidence, the authenticity of the data before it was written to the chain and the reliability of the technical platform. Where a party offers AI-generated content as evidence of infringement, the court considers the design of the prompt and its effect on the output, the degree of similarity between the generated content and the asserted work, the consistency of repeated testing, and factors such as model training, algorithm design, and content filtering mechanisms.

Verification and disclosure obligations for AI-generated filings

The first half of Article (19) addresses fraudulent litigation: where a party obtains false evidence through human intervention or misleading means — deleting or altering the synthetic content label, specific instruction inputs, selective presentation of results, adversarial interference — and fabricates the basic facts of a civil case to bring fraudulent litigation, the court dismisses the claim and, depending on severity, imposes fines or detention, with criminal liability where a crime is constituted. Where a litigation participant or other person uses AI to fabricate evidence and obstructs adjudication, Article 114 of the Civil Procedure Law applies.

The second half binds counsel directly: where a litigation participant submits pleadings, case research reports, or similar materials that were generated using AI, "before submission to the court they shall carefully verify the authenticity and accuracy of the laws, judicial interpretations, cases, and other content; at the time of submission they shall explain the use of AI assistance; and they shall bear responsibility according to law for the authenticity and accuracy of that content." Three obligations in parallel: verify before submission, disclose at submission, and answer for authenticity and accuracy.

The Court's answers explain the provenance: "people's courts have already identified multiple instances of AI-generated false cases, and the parties and counsel had not fully verified them before submission to the court," with entries already in the people's courts case database. The Court also noted that "foreign judiciaries are equally attentive," with some countries issuing dedicated guidance emphasizing counsel's responsibility to ensure the accuracy of materials submitted. The same problem has produced sanctions repeatedly in U.S. federal courts — see our earlier note, Fabricating Case Law With AI: Dismissal, Sanctions, and Professional Consequences.

Criminal liability

Article (20) enumerates circumstances in which using AI to commit fraud, insult, defamation, damage to business reputation or product reputation, infringement of citizens' personal information, unlawful acquisition of computer system data, or the production, sale, and dissemination of obscene materials constitutes a crime. It separately addresses one situation: causing a traffic accident by using privately installed accessories to evade assisted-driving system monitoring after activating the assisted-driving function, with criminal liability where a crime is constituted.

V. Working Mechanisms (Articles 21–24)

Part Five provides for four things: using diversified dispute resolution mechanisms and establishing coordination with AI industry regulators, industry and professional mediation organizations, and academics; standardizing the use of elevated jurisdiction for cases involving major interests, difficult and complex new types, rule-setting significance, or the need to unify the application of law, and strengthening case guidance through the people's courts case database; strengthening coordination with cyberspace, public security, procuratorial, and market regulation authorities; and advancing exchange and cooperation in foreign-related adjudication in the AI field and properly adjudicating cross-border AI and data disputes.

Observations

The Opinion's rule density is concentrated in Parts Two and Three, where the same structure recurs: liability is placed on what a particular party is able to prove. The developer must be able to produce the training data source, training process records, and model operating mode; the open-source party must be able to produce its public explanation of function and security risks; the producer and seller must be able to produce their explanation and warnings about application scenarios and inherent limitations; the data controller must be able to produce complete event records. Whether those materials were created before the dispute arose, and whether they can be retrieved, will directly determine the evidentiary outcome.

Two points on scope are also worth noting. Article (9) confines product liability to products with a physical carrier, excluding purely software-form AI services, which must return to fault liability. The second half of Article (19) runs the other way: it is not limited to AI-related cases but imposes verification and disclosure obligations on the use of AI to assist in preparing litigation materials in all cases.

---

The full text appears in our republication. Original sources: Supreme People's Court website, "SPC Issues the Opinion on Adjudicating Disputes Involving Artificial Intelligence in Accordance with Law" and "Responsible Officials of Relevant SPC Departments Answer Press Questions on the Opinion" (both published September 7, 2026). Citations to the Opinion are taken from its text; statements of the Court from the two release documents are paraphrased and quoted as appropriate.

This article is a review of published judicial documents. It does not constitute legal advice and does not create an attorney-client relationship. Specific matters should be addressed with counsel in light of the particular circumstances.