Skip to content

China’s Supreme People’s Court Issues Its First Comprehensive Guidance on AI-Related Disputes: An IP Enforcement Perspective

China IP Brief | Dispatches from the Frontiers of IP Law in China  

I. Introduction

On September 7, 2026, the Supreme People’s Court (SPC) of the People’s Republic of China (PRC) issued the Opinion on Adjudicating AI-Related Disputes in Accordance with the Law (Fa Fa [2026] No. 10). Comprising 24 articles organized into five sections, the Opinion is the first judicial policy document issued by the PRC’s highest judicial authority focused on the adjudication of AI-related disputes.

An unofficial AI-generated and human-edited English translation of the Opinion is attached here.

The widespread use of AI in content creation, research and development, and commercial operations has raised novel questions regarding the application of existing law. Drawing on domestic and international jurisprudence, industry experience and academic research, the Opinion addresses key issues arising in the adjudication of AI-related disputes in the PRC. Building on the PRC Civil Code, the Personal Information Protection Law (PIPL), the PRC Copyright Law, the PRC Anti-Unfair Competition Law (AUCL) and other applicable legislation, the Opinion outlines adjudication guidelines based on principles for which a working consensus has emerged in PRC judicial practice.

The Opinion is broad in scope, and it seeks to balance AI innovation with the need to address issues involving the allocation of liability, evidentiary burdens and judicial procedure in AI-related disputes. This alert analyzes the Opinion from the perspectives of IP disputes and enforcement strategy, corporate AI compliance and litigation practice.

II. General Principles of Liability in AI-Related Cases and the Liability of AI Service Providers

1. General Basis of Liability in AI-Related Cases (Article 3)

Article 3 of the Opinion sets out the general basis for liability in AI-related tort disputes. Where a specific statute provides for strict liability (i.e., no-fault liability) or presumed-fault liability, that statute will govern. In the absence of such a provision, courts should apply the general fault-based liability principle under Article 1165(1) of the PRC Civil Code. For example, harm caused by AI products, such as embodied AI systems, may be subject to strict liability under the PRC Civil Code and the PRC Product Quality Law. Harm arising from the use of AI technology to process personal information may be subject to presumed-fault liability under the PIPL. Other forms of generative AI-related tortious conduct generally remain subject to ordinary fault-based liability unless otherwise provided by law.

Article 3 adopts a context-specific approach to fault. In determining whether and to what extent a party is at fault, courts should consider six factors: (i) the AI system’s use case; (ii) the degree of autonomy; (iii) the transparency of the relevant technology and information; (iv) the potential risks and scope of impact; (v) the preventive and mitigating measures taken by AI developers, providers and other relevant parties, as well as the technical feasibility of such measures; and (vi) the AI user’s ability to foresee and control the harm that may result from tortious conduct involving AI technology.

The first five factors largely concern the duties of AI developers and service providers, while the sixth focuses on the conduct and responsibility of users. The importance of operational safeguards is illustrated by China’s first reported “AI hallucination” infringement case ((2025) Zhe 0192 Min Chu No. 18143). In this case, the court found that the platform was not at fault after considering its defense evidence of risk warnings, product disclosures, and risk prevention and control records.

2. Liability of Generative AI Service Providers (Article 7)

Generative AI service providers connect upstream technology developers with downstream users of AI-generated content and generally exercise control over content-generation mechanisms, data flows and channels of dissemination. They are therefore central actors in AI-related infringement cases.

Article 7 of the Opinion addresses the liability of generative AI service providers and, for the first time at the judicial policy level, confirms that the traditional internet platform “safe harbor rules” may be applied by analogy to generative AI services. Although Article 7 principally concerns infringement of personality rights, it is likely to be instructive for the handling of cases involving IP infringements as well.

Article 7 provides that where AI-generated content infringes personality rights and interests, a service provider that fails to take “necessary measures” in a timely manner after receiving notice from the rights holder may be liable for resulting harm, and a user who intentionally uses infringing prompts to induce the generation of unlawful content may also be liable. The previous debate centered on whether generative AI substantively “participates” in content generation and whether the absence of an infringing user, such as in AI hallucination scenarios, fits the traditional tripartite structure of online infringement. Article 7 signals that the SPC has adapted the traditional safe harbor for online infringement—the “notice and takedown” model—to generative AI service providers. In this context, however, the framework is better understood as “notice and necessary measures” rather than conventional notice and takedown, because AI-generated content may be delivered privately on a one-to-one basis, leaving no publicly accessible content or link to remove or disable.

In light of the technical features of AI-generated content, the Opinion explains that a rights holder may provide an effective notice by reasonably demonstrating the generation of infringing content through simulated prompt inputs, technical testing or similar means, and that the notification threshold should not be set unduly high. Moreover, the scope of “necessary measures” extends beyond deleting or blocking content and disabling links. Depending on the circumstances, such measures may include (i) preventing the further generation of the infringing content, (ii) blocking relevant prompts or generation instructions, and (iii) warning users or restricting or suspending their access to the service. These measures are prospective and prevention oriented, reflecting the distinctive technical features of generative AI services.

In addition, Article 7 calls for “prudence” in determining the liability of generative AI service providers, reflecting an approach that seeks to balance the promotion of AI innovation against the protection of rights holders while avoiding the imposition of liability burdens that exceed the current stage of technological development. Notably, the final Opinion removed language from an earlier draft expressly referring to a “knew or should have known” standard, which is associated with the “red flag” rule. This omission suggests a deliberate effort to avoid imposing unduly burdensome monitoring obligations on the AI industry at its current stage of development.

Key Practical Considerations:

For generative AI service providers, express noninfringement undertakings should be incorporated into user agreements as a foundational measure for invoking safe harbor protection. Providers should also implement baseline content review procedures and maintain accessible infringement reporting channels. Upon receiving notice, providers should apply graduated measures to repeat infringers, such as warnings, functionality restrictions and, ultimately, service suspension and retain complete records showing when the notice was received, how it was reviewed, what measures were taken and when those measures were implemented. Enhanced scrutiny should be applied to content involving well-known works, trademarks, trade names and other high-profile subject matter. Product lines that edit, select, curate or recommend content should undergo separate risk assessments and, where appropriate, enhanced review.

For rights holders, an infringement notice should contain preliminary evidence of infringement and accurate identifying information for the notifying party. When preserving evidence through simulated testing, rights holders should record the date and time of each test, the model and model version used, all inputs and outputs, and the complete interaction context. Where appropriate, repeated testing should be conducted to demonstrate the consistency and reproducibility of the challenged outputs. Rights holders should also carefully design and document their interactions with the AI system to avoid allegations that they intentionally induced the generation of infringing content.

III. Liability Boundaries for AI-Related IP Infringements (Articles 12 – 16)

1. Copyright Infringement Involving AI-Generated Content: Allocation of Liability and Burdens of Proof (Article 12)

Article 12 of the Opinion addresses the allocation of liability among AI developers, service providers and users in cases involving allegedly infringing AI-generated content.

It adapts the safe harbor principles under Articles 1195 and 1197 of the PRC Civil Code to the AI context and systematizes the experience of the Chinese courts over the past three years.

In the Guangzhou Ultraman case ((2024) Yue 0192 Min Chu No. 113), a platform that offered image generation services through an API connection to a third-party model was found directly liable because of its substantial participation in the generation of the infringing content. In the Hangzhou Ultraman case ((2024) Zhe 0192 Min Chu No. 1587, affirmed on appeal), a platform that provided model access and storage services was held liable for contributory (aiding) infringement for failing to meet its duty of care. By contrast, in the Shanghai Medusa case ((2025) Hu 0116 Min Chu No. 2354, affirmed on appeal), a comparable platform avoided liability because it had implemented appropriate ex ante review and infringement reporting mechanisms and responded promptly after receiving notice.

Article 12 directs courts to consider six principal factors: (i) the type of AI service, (ii) industry characteristics, (iii) the source of training data, (iv) each party’s degree of participation, (v) the “necessary measures” taken and (vi) the commercial benefit obtained. These factors reflect a broader principle—that liability should correspond to a party’s ability to control the relevant risk and the scope of its duty of care. A party’s degree of participation in generating or disseminating the allegedly infringing content is likely to be particularly important, as illustrated by the different outcomes in the Guangzhou and Hangzhou Ultraman cases.

Article 12 also addresses the evidentiary obligations of the principal participants in the AI supply chain. Where an AI developer raises a noninfringement defense, the court should order it to provide supporting materials—i.e., (i) the sources of training data, (ii) records of the training process, (iii) the model’s operating mode and (iv) the scientific or technical basis supporting its position. Where a rights holder alleges that a generative AI service provider has used algorithmic technology to infringe its copyright, the rights holder must provide supporting evidence. Where an AI service user “knew or should have known” of a preexisting work, used AI to generate a work substantially similar to that work and lacks a reasonable defense, the people’s court should uphold the copyright owner’s claim for infringement.

At the press conference accompanying the release of the Opinion, SPC officials further explained that a rights holder may satisfy its initial burden of proof against an AI developer or generative AI service provider by establishing that the challenged content was generated by the AI system in question and that the content is substantially similar to the copyrighted work. Where a claim is brought against an AI user, evidence that the user had a reasonable possibility of access to the preexisting work may give rise to a presumption that the user knew of the work. The burden may then shift to the user to rebut that inference. A user may therefore be held liable if the user knew or should have known of the preexisting work, used AI to generate a substantially similar work and lacks a reasonable defense. These evidentiary rules materially lower the barriers to enforcement for rights holders confronting the algorithmic “black box.”

Key Practical Considerations:

For rights holders and AI developers, rights holders asserting claims against AI users should preserve evidence of the publication date of the preexisting work, dissemination and market exposure to establish a reasonable possibility of access and support a presumption that the alleged infringer knew or should have known of the work. AI developers, in turn, should establish a traceability system covering the entire data and model life cycle—from sourcing and cleaning through training and iteration. Relevant records should include inventories of training data sources and associated licenses, data cleaning and processing records, model version and iteration logs, and records of risk testing and human intervention. These materials may be critical to substantiating a noninfringement defense.

For generative AI service providers and users, providers should clearly set out IP compliance requirements in their user agreements, conduct appropriate prepublication reviews of user-uploaded training materials and models intended for release, and retain complete records of their review and remedial actions. AI users should avoid entering complete preexisting works into prompts for the purpose of inducing infringing outputs and should conduct a reasonable clearance review before commercially exploiting AI-generated content. 

2. Open-Source Software: A New Liability Exemption (Article 13)

Article 13 of the Opinion introduces a specific framework for evaluating the liability of developers and providers of open-source software used in AI development. In determining whether an appropriate exemption from liability applies, courts should consider (i) the type of open-source license, (ii) the specific scope of the rights limitations, (iii) the security and compliance measures taken and (iv) the extent of information disclosure.

Article 13 further provides that where a developer or provider makes certain code modules required for AI software development available free of charge on an open-source basis and publicly discloses their functions and security risks, a court may find that the developer or provider is not liable for infringement arising from a third party’s use of those modules. This framework reflects the Opinion’s broader focus on the relevant party’s actual control over the risk of infringement and degree of participation in the allegedly infringing conduct. Liability should not automatically extend upstream to an open-source contributor merely because its code was incorporated into a downstream product.

Key Practical Considerations:

Open-source developers and providers should select appropriate licenses; clearly define applicable rights restrictions; disclose each module’s functions, scope of use and known security risks; and retain security review and vulnerability testing records to support any liability exemption.

Downstream companies should maintain component inventories and license review checklists, conduct security assessments and contractually address data inputs, control of high-risk functions and investigative assistance. Reliance on upstream open-source components does not relieve downstream companies of their own duty of care.

3. Patents and Technology Contracts (Articles 14 – 15)

Article 14 confirms that an AI-related invention may qualify as patentable subject matter if it (i) employs technical means consistent with the laws of nature, (ii) solves a technical problem and (iii) produces a technical effect consistent with the laws of nature. It also confirms that an individual who uses AI to complete an invention may be recognized as the inventor if that individual made a creative contribution to the substantive features of the invention. In addition, an AI-related patent satisfies the “sufficiency of disclosure” requirement if its specification describes the technical solution in sufficient detail to enable implementation by a person skilled in the art. As AI becomes more deeply integrated into research and development, whether a named human inventor made the requisite “creative contribution” is likely to become a key issue in patent validity and ownership disputes involving AI-assisted inventions.

Article 15 provides that liability for breach of AI-related technology contracts is determined on the basis of the contract, with full consideration of the characteristics of AI technology R&D and whether the developer exercised reasonable efforts.

Key Practical Considerations:

For companies using AI in the R&D process, contemporaneous records should document human creative contributions, including prompt design and revision, parameter adjustments, the selection and further processing of outputs, and iterations of technical solutions and design documents. The greater AI’s role in the R&D process, the more important such records become. Companies should also implement internal disclosure and assessment procedures for AI-assisted inventions and evaluate the extent of human contribution before filing patent applications.

For parties to AI-related technology contracts, agreements should clearly address (i) milestones and acceptance criteria; (ii) the means of demonstrating compliance with reasonable efforts obligations, such as the periodic submission and review of R&D records; and (iii) the allocation of risks and costs if development does not achieve the expected results.

4. A Layered Framework for Protecting Data Rights and Interests (Article 16)

Article 16 affirms that AI-related data rights and interests are protected by law and establishes a layered framework based on the nature of the data or dataset. A dataset qualifying as a compilation or another type of copyrightable work may be protected under the PRC Copyright Law. Data satisfying the requirements for trade secret protection may be protected under the trade secret provisions of the PRC AUCL. Even where data does not qualify as a trade secret, acquiring or lawfully using data held by another business operator through improper means may result in liability under Article 13 of the AUCL.

The Opinion also expressly prohibits two categories of conduct: (i) using data or algorithmic technologies to enter into monopoly agreements or abuse a dominant market position; and (ii) compromising the secure operation of AI systems through fabricated or disruptive data, malicious data labeling, adversarial sample attacks or similar technical means.

Key Practical Considerations:

For data holders, the appropriate protection strategy and supporting evidentiary record should be established for each category of data. Copyright claims require evidence of originality in the selection or arrangement of the dataset; trade secret claims require demonstrable confidentiality measures, such as tiered access controls, nondisclosure agreements and access logs; and claims under Article 13 of the AUCL require records of substantial investment in acquiring, organizing, processing and maintaining the data. Data holders should also ensure that data supply and collaboration agreements address the transmission of individuals’ objections and changes in authorization throughout the data supply chain and map data batches to authorized uses, allowing potential claims to be linked to specific processing stages. In addition, they should review data- or algorithm-based exclusivity arrangements for antitrust risk and establish security and evidence-preservation protocols for fabricated or disruptive data, malicious labeling  and adversarial sample attacks. 

IV. Procedural Rules (Articles 17 – 19)

Article 17 directs courts to strengthen procedural guidance and clarification, investigate and collect evidence ex officio where appropriate and promptly review applications for evidence preservation. Parties may also apply for court-assisted evidence collection. Critically, Article 17 establishes an adverse inference rule for spoliation or obstruction of evidence: If a party controlling documentary or electronic evidence refuses to produce it without justification, the court may accept the opposing party’s allegations concerning the contents of that evidence where those contents would be unfavorable to the controlling party. This mechanism is likely to make Article 12’s requirement that AI developers produce information concerning training data, training processes and model operation more effective in practice. For technically complex issues, courts may draw on lay assessors, court-appointed appraisers, party-retained expert assistants, technical investigation officers and other technical resources.

Article 18 establishes differentiated standards for reviewing AI-related evidence. For big-data analysis reports, courts should examine (i) the source of the underlying data, (ii) the applicable data cleaning rules and (iii) the scientific reliability of the analytical methodology. For blockchain-preserved evidence, courts should examine (i) the authenticity of the data before it was recorded on-chain and (ii) the reliability of the relevant technical platform. Where AI-generated content is introduced as evidence of infringement, courts should consider (i) the design of the prompts and their influence on the outputs, (ii) the degree of similarity between the generated content and the protected work, (iii) the consistency of repeated tests and (iv) the model’s training and filtering mechanisms. Accordingly, a persuasive evidentiary record should include repeated testing with complete records of time, model version, prompts, contextual inputs, outputs and relevant system settings so as to build an evidentiary chain capable of withstanding cross-examination.

Article 19 regulates the use of AI for fraudulent litigation and evidence fabrication and establishes a duty to verify and disclose AI-generated litigation materials. In a 2025 Ren X Lu case included in the SPC’s official case database ((2024) Jing 0112 Min Chu No. 19067), counsel cited an AI-generated “case” that used a genuine case number but contained entirely inaccurate substantive information. The court formally admonished counsel for this conduct. All AI-assisted litigation materials should therefore be subject to independent professional review and verification against authoritative sources before submission.

Key Practical Considerations:

The Opinion indicates that Chinese courts will place increasing emphasis on rigorous fact-finding, evidentiary reliability and the integrity of judicial proceedings in AI-related cases. Courts are expected to scrutinize the authenticity, completeness and verifiability of AI-generated content, electronic data and algorithm-related evidence, giving particular attention to training data sources, content-generation processes, prompt design, and model architecture and operation. The Opinion also directs courts to guard against the use of AI to fabricate evidence, generate false materials at scale or otherwise mislead the court.

International rights holders seeking to enforce copyright, trademark, patent or trade secret rights in China should therefore take steps to preserve evidence of infringement at an early stage. Relevant materials may include AI-generated outputs, testing protocols and results, technical analysis reports and evidence preserved through blockchain-based systems. Rights holders should also establish a complete and well-documented chain of evidence and develop technical arguments addressing the accused model’s training data sources, output-generation mechanisms and connection to the alleged infringement. These measures can strengthen the evidentiary record and improve the effectiveness of enforcement proceedings.

V. Conclusion

The Opinion reflects China’s shift from case-by-case experimentation toward a more systematic framework for adjudicating AI-related disputes. For generative AI service providers, the Opinion establishes fault-based liability as the baseline, with the “notice and takedown” mechanism—and other necessary measures—playing a central role. Article 7 extends the safe harbor framework traditionally applied to online infringement to generative AI services. More broadly, the Opinion reflects a cautious approach to determining provider liability. Providers are not expected to conduct ex ante review of the vast volume of content generated through their services. Once they receive a valid notice, however, they must take timely and appropriate action within the scope of their technical and managerial capabilities. At the same time, the Opinion preserves flexibility to impose greater responsibility on providers whose business models afford them greater control over, or deeper involvement in, the generation or dissemination of content.

With respect to IP protection, Article 12 establishes a liability allocation framework under which responsibility is calibrated to a party’s control capabilities and duty of care. Its burden-of-proof rules also alleviate, to some extent, the evidentiary difficulties rights holders encounter when confronting the black box of algorithmic systems. The exemption for certain open-source activities, the rules governing patent eligibility and inventorship for AI-related inventions, and the tiered protection of data-related rights likewise provide participants across the AI value chain with clearer expectations regarding permissible conduct and potential liability.

Articles 17 through 19, which address adverse inferences arising from spoliation or obstruction of evidence, differentiated standards for reviewing evidence and the obligation to verify AI-generated litigation materials, further enhance the practical enforceability of these substantive rules. Taken as a whole, the Opinion does not fundamentally reconstruct China’s existing IP regime. Rather, it systematically consolidates and codifies, within the current legal framework, jurisprudential principles that have gradually emerged from AI-related cases in recent years.

For foreign IP rights holders, the Opinion signals that China is developing a more tailored judicial framework for disputes involving AI model training, content generation, data use and proof of infringement. Chinese courts continue to take a measured approach to liability, seeking to balance technological innovation against the protection of IP and other proprietary interests. At the same time, the Opinion may lower practical barriers to enforcement in AI-related disputes by strengthening the protection of data-related rights, refining evidentiary and burden-of-proof rules and enhancing procedural safeguards.

Against this backdrop, rights holders should take a proactive approach to enforcement. Whether a dispute involves unauthorized model training, AI-generated content that infringes copyrights or trademarks, or claims concerning trade secrets, data-related rights or technology agreements, rights holders should prioritize technical investigation, implement robust evidence-preservation protocols and engage Chinese counsel and qualified technical experts at an early stage.

As the frequency of AI-related litigation continues to grow, the Opinion is likely to serve as an important reference point for Chinese courts and as a meaningful indicator of China’s broader AI and IP policy direction in the coming years. International companies developing, deploying or otherwise conducting AI-related business in China should closely monitor how Chinese courts interpret and apply the Opinion in practice.