China’s Supreme People’s Court Clarifies Legal Rules for AI-Related Disputes

I. The Opinions and Their Practical Significance

China’s Supreme People’s Court (the “SPC”) has issued its first dedicated guidance on the adjudication of AI-related disputes, setting out how existing legal rules should apply across tort, intellectual property and litigation procedure while leaving several important questions unresolved. The Opinions on the Adjudication of AI-Related Disputes in Accordance with Law (the “Opinions”), issued on 7 September 2026, contain 24 articles covering AI-related torts, intellectual property, litigation procedure and related judicial mechanisms.

The Opinions are an SPC adjudicative guidance document, not a judicial interpretation. Under the SPC’s rules on legal citations and judicial reasoning, civil judgments should cite applicable laws, legislative interpretations and judicial interpretations; applicable administrative regulations and certain local regulations may also be cited directly. Other valid normative documents may be used as part of the court’s reasoning. Courts will therefore continue to decide AI disputes under existing legislation, including the Civil Code, the Personal Information Protection Law (the “PIPL”), the Copyright Law and the Product Quality Law. Even so, the Opinions are likely to have significant practical influence on how Chinese courts approach novel AI disputes, apply existing law and assess relevant factors.

The Opinions also bring into civil adjudication many of the same technical and governance considerations that already feature in China’s administrative regulation of algorithms, deep synthesis and generative AI. Existing regulatory rules focus on matters such as algorithm and model transparency, risk assessment, training-data sources, content controls, complaint handling and log retention. The Opinions use similar considerations in civil cases: Article 3 makes transparency, risk, preventive measures and technical feasibility relevant to fault; Article 7 focuses on a service provider’s response after receiving an infringement notice; and the intellectual-property and evidence provisions examine training data, training processes, model operation and content-filtering mechanisms. The same technical and governance facts may therefore become relevant both to regulatory compliance and to civil litigation.

II. How Existing Legal Rules Apply to AI

A. Fault Remains the Default Where No Special Liability Rule Applies

The Opinions retain fault-based liability as the default for AI-related torts where the law does not expressly impose strict liability or presumed-fault liability. Article 3 builds on the general fault principle under Article 1165(1) of the Civil Code and directs courts to consider factors such as the use case, the system’s degree of autonomy, technical and information transparency, the nature and reach of potential risks, and the ability of developers, service providers and users to prevent or control harm.

For businesses, what risks were known, what safeguards were technically available and how the business actually responded may all affect the court’s assessment of fault. The degree of system autonomy is one factor, while knowledge of the relevant risks, practical control over those risks and the proportionality of the measures taken may also carry significant weight.

The Opinions apply existing civil-law rules across a range of AI-related harms. Articles 4, 5 and 10 address practices including AI face swaps and voice cloning, doxxing and data-driven price discrimination through existing rules on personality rights, consumer protection and tort liability, while Article 8 addresses injunctive relief against imminent or ongoing infringements of personality rights. The legal analysis therefore continues to depend on the right affected, the conduct involved and the responsibility attributable to the relevant party.

B. Software Services and AI-Enabled Products Follow Different Liability Paths

The applicable liability framework depends in part on how an AI system causes harm. Where a generative AI service infringes personality rights through text, images or audio, fault-based tort liability and the relevant notice-and-action mechanism will generally be the primary routes. Where AI is embedded in a physical product, such as a robot or autonomous vehicle, and a product defect causes personal injury or property damage, product liability may apply. The form of the technology, the mechanism of harm and the type of loss therefore help determine the appropriate legal framework.

In its accompanying responses to press questions, the SPC stated that AI services that are not embodied in a physical product fall outside the traditional product-liability regime. Article 9 accordingly focuses on AI-enabled physical products that fall within the definition of a “product” under the Product Quality Law. It also identifies factors relevant to defect analysis, including the nature and intended use, autonomous learning and updating capabilities, the user’s degree of control, applicable standards, and risk disclosures and warnings.

Article 11 illustrates how several existing liability regimes may operate together in autonomous-driving accidents. Depending on the cause of the accident, traffic-accident liability, product liability arising from a vehicle defect and rules allocating liability for separate tortious acts may each be relevant.

C. Notice-and-Action Duties Extend to Future AI Outputs

Article 7 adapts the notice-and-action framework for online infringement to the dynamic nature of generative AI. On a conventional online service, a necessary measure will often target an existing post or link. For a generative AI service, effective action may also need to address future generation of the relevant infringing content.

Where automatically generated content infringes a person’s reputation, privacy or other personality rights, the rights holder may submit a notice containing preliminary evidence and their true identity. If the service provider fails to take necessary measures in a timely manner, such as stopping the generation of the infringing content, it is liable for the resulting harm.

Article 7 separately addresses malicious prompting by users. A user who maliciously induces a model to generate infringing content through infringing prompts or similar means, and thereby causes harm, is liable for that harm. If, after receiving notice, the service provider fails to take necessary measures such as stopping generation of the infringing content or blocking the relevant prompts, a rights holder may seek liability under Article 1195 of the Civil Code.

The SPC’s accompanying explanation describes this approach as an application by analogy of the notice-and-action mechanism used for network service providers. The functional similarity is that both conventional network services and generative AI services may be able to limit continuing harm after receiving notice. In the generative-AI context, however, the response may extend beyond dealing with existing content to controlling future generation. Complaint channels, notice verification, model responses, prompt blocking, the effectiveness of remedial measures and records of the response may all become relevant to liability. The Opinions leave the broader private-law classification of generative AI service providers open.

D. Publicly Available Personal Information May Be Used for Training Within Limits

Article 6 applies the rule in Article 27 of the PIPL on publicly available personal information to model training and gives courts more concrete factors for determining whether processing remains within a “reasonable scope.” Where information has been made public by the individual or otherwise lawfully disclosed, processing it for model training within a reasonable scope will generally not be treated as an infringement of personal-information rights if the individual has not expressly objected. Consent remains required where the processing would have a significant impact on the individual’s rights and interests.

In assessing whether the processing remains within a reasonable scope, courts should consider the necessity and appropriateness of the training purpose in light of the model’s function, the type and sensitivity of the information, the potential impact on the individual, the context in which the information was originally disclosed and the individual’s reasonable expectations regarding later use. The Opinions therefore provide a more concrete framework for applying an existing PIPL rule to model training.

E. Output-Side Copyright Liability Is Clearer Than Training-Side Legality

Article 12 focuses on liability where AI-generated output infringes another person’s copyright. Courts should consider the type of AI service, industry characteristics, the source of training data, each party’s degree of involvement, the necessary measures taken and any profits derived.

If a developer raises a non-infringement defence, the court is directed to require it to provide supporting materials concerning the sources of training data, records of the training process, the model’s operating mode and the relevant scientific basis. This brings model operation, training materials and the parties’ respective involvement directly into the assessment of infringement, liability and evidence.

Article 12 addresses infringement at the output stage. It leaves open two separate questions: the legal treatment of using copyrighted works for model training without authorization, and when AI-generated content itself may qualify for copyright protection. Those issues are discussed further below.

F. Open-Source Liability Turns on Control, Involvement and Disclosure

Article 13 provides a separate set of considerations for open-source AI. In assessing the liability of open-source software developers and providers, as well as downstream developers and providers, courts should consider the type and restrictions of the applicable open-source licence, safety and compliance measures, and the extent of information disclosed.

Where a developer or provider makes certain code modules available free of charge on an open-source basis and publicly explains their functions and safety risks, a court may, depending on the circumstances, find that the developer or provider is not liable for infringement arising from downstream use. The rule places particular weight on each party’s actual control, level of involvement and risk disclosures, and provides a more concrete basis for distinguishing the responsibility of foundational code providers from that of downstream application providers.

Beyond open-source liability, Articles 14–16 address other commercially significant issues. Article 14 covers patent eligibility, inventorship and sufficient disclosure for AI-related inventions. Article 15 directs courts to assess AI technology-contract breaches by reference to the contract, the nature of AI R&D and whether the developer used reasonable efforts, and Article 16 protects lawfully acquired data interests while applying copyright, trade-secret, unfair-competition and antitrust rules to data use and conduct affecting AI-system security.

G. Governance Records May Become Litigation Evidence

Technical facts and governance records may become increasingly important in determining liability and proving a case. Read together, Articles 3, 7, 12, 17 and 18 indicate that materials concerning training-data sources, training processes, model operating modes, prompts, filtering mechanisms, repeat-test results and complaint-response records may be relevant to establishing infringement, a party’s ability to control the system, the measures it took and the availability of particular defences. Article 17 also allows a court to draw an adverse conclusion where a party controlling documentary or electronic evidence refuses to produce it without proper justification and the opposing party asserts that the evidence would be unfavorable to the controlling party.

Records created through ordinary AI governance may therefore later become litigation evidence. A company’s ability to preserve, retrieve and explain records relating to training data, model versions, risk testing, content filtering and complaint handling may directly affect its ability to establish its case or defend a claim.

Article 19 also addresses the use of AI in litigation. Participants submitting AI-generated litigation documents, case-search reports or similar materials must verify the authenticity and accuracy of cited laws, judicial interpretations and cases, disclose AI assistance when submitting the materials, and remain responsible for their accuracy. It also provides for sanctions where AI is used to fabricate evidence or pursue false litigation.

III. Key Copyright Questions Remain Open

A. Copyright Issues in the Use of Works for Model Training and AI-Generated Content Remain Unsettled

The Opinions leave two central copyright questions unresolved. In its responses to press questions, the SPC acknowledged substantial disagreement over whether AI-generated content may itself qualify for copyright protection and the legal treatment of using another person’s copyrighted works to train a large model without authorization. The first concerns rights in generated content; the second concerns the legal boundaries of model training. Both have direct implications for access to training data, model development and the commercialization of AI-generated content.

The Opinions do, however, provide rules for copyright infringement at the output stage. Article 12 addresses how liability may be allocated among developers, service providers and users when AI-generated content infringes another person’s copyright, as well as how training-related materials may be used in assessing infringement and defences. Its focus is therefore on liability and evidence once allegedly infringing output has been generated. Questions concerning the legal treatment of using copyrighted works for model training without authorization and the copyright status of AI-generated content remain open and should be assessed separately from output-side infringement risk.

B. Future Cases and SPC Case Guidance as Key Near-Term Indicators

Article 22 leaves room for unresolved issues to develop through judicial practice. AI-related disputes involving major interests, difficult or complex new issues, broader rule-setting significance or a need for consistent legal standards are to be elevated for adjudication by a higher-level court. The Opinions also call for fuller use of the People’s Court Case Database to strengthen case guidance and improve consistency in adjudication.

Until more systematic or issue-specific rules emerge, cases raising issues of broader significance, cases heard by higher-level courts and relevant entries in the People’s Court Case Database will be important near-term indicators of how Chinese courts approach unsettled issues, including the use of copyrighted works in model training and the copyrightability of AI-generated content. As judicial experience accumulates and standards become more settled, these issues may be clarified further through subsequent judicial practice and legal developments.

IV. Key Takeaways

1. The Opinions are adjudicative guidance, not a judicial interpretation. Courts must continue to decide AI-related disputes under existing law, but the Opinions are likely to shape judicial methodology, legal analysis and the factors considered in practice.
2. There is no single liability regime for AI. The applicable rules depend on how the technology is deployed, how the harm occurs and the role of each party. Software-based services, AI-enabled physical products, personal-information processing and copyright disputes may therefore follow different legal paths.
3. The Opinions clarify output-side copyright liability but leave two major questions open. They do not determine the legal treatment of using copyrighted works for model training without authorization or when AI-generated content may itself qualify for copyright protection.
4. AI governance and litigation evidence are becoming increasingly connected. Technical and governance records created in the ordinary course of business may directly affect liability, proof and available defences in AI-related disputes.