The Supreme People’s Court issues opinions on the trial of AI-related disputes
Background
In recent years, generative artificial intelligence (hereinafter referred to as GenAI) has been widely applied in scenarios such as marketing, customer service, research and development, and driving. The development of related products and services has also given rise to various new types of legal disputes, while courts have so far lacked unified standards for adjudicating AI-related disputes and China has not adopted a comprehensive AI law. Against this background, on 7 September 2026, the Supreme People’s Court issued Opinions on the Lawful Trial of AI-Related Disputes (hereinafter referred to as Opinions), providing directional guidance for the adjudication of AI-related cases.
This article, by reference to publicly available judgments and rulings relating to artificial intelligence, will focus on Parts II (adjudication of AI-related tort disputes) and III (adjudication of AI-related IP disputes) of the Opinions and share our observations on the key points of these sections.
II. AI-related tort cases
(I) Accurately identifying the applicable attribution principles for tort liability in AI-related cases
Civil Code, the Personal Information Protection Law, the Product Quality Law, the Road Traffic Safety Law and other relevant laws and regulations. The applicable liability rules should therefore be determined in accordance with the relevant statutory framework. At the same time, where laws and regulations do not expressly provide for no-fault liability or presumed-fault liability, the fault-based liability principle under Article 1165(1) of the Civil Code – namely, that “where an actor, through fault, infringes upon the civil rights and interests of another person and causes damage, the actor shall bear tort liability” – should serve as the standard for determining whether the actor should be held liable in tort.
In assessing whether an actor is at fault and the degree of such fault in a specific case, three aspects should be considered comprehensively:
- Application of AI technology: The specific application scenario of the AI, its degree of autonomy, the transparency of the technology and information, potential risks and the scope of its impact.
- Safeguard measures objectively adopted by developers, providers and others: The measures taken by AI developers, providers and other relevant parties to prevent and reduce AI-related torts, as well as the technical feasibility of such measures.
- The user’s subjective state of mind: The AI user’s ability to foresee and control the damage that may be caused by tortious conduct carried out through AI technology.
[Case for Example] In the personality rights dispute between Yin and a Beijing-based intelligent technology company and others, the defendants used AI technology to process audio materials recorded by the plaintiff as a voice actor and generated the text-to-speech product at issue for commercial gain. After an in-court inspection, the Beijing Internet Court confirmed that the AI-generated voice was highly consistent with the plaintiff’s voice timbre, intonation, pronunciation style and other characteristics. Although Defendant 2 presented a data authorisation letter to Defendant 3, the letter was not signed by the plaintiff personally. As none of the defendants had obtained the plaintiff’s informed consent, the court held that certain defendants had infringed the plaintiff’s voice-related rights and interests.
(II) Regulating, in accordance with law, acts that use generative AI to infringe personality rights and interests, such as “AI face-swapping and voice imitation” and “AI resurrection of the deceased”
According to the Opinions, unless prior consent has been obtained from the natural person concerned or the law provides otherwise, no party may use AI to process a natural person’s name, likeness (which is not limited to facial features, but can include other characteristics as long as the public can identify a specific individual from such characteristics) or other attributes to generate, use or publish a virtual digital image by which that natural person can be identified; use a natural person’s voice as training material to imitate and synthesise a human voice; manipulate a virtual digital image or voice by which a specific individual can be identified to carry out improper acts or make false statements; or create or use a virtual digital image of a deceased person in a manner that infringes upon the deceased person’s name, likeness, reputation or other rights and interests.
[Case for Example] In the personality rights dispute between Yin and a Beijing-based intelligent technology company and others, the defendants used AI technology to process audio materials recorded by the plaintiff as a voice actor and generated the text-to-speech product at issue for commercial gain. After an in-court inspection, the Beijing Internet Court confirmed that the AI-generated voice was highly consistent with the plaintiff’s voice timbre, intonation, pronunciation style and other characteristics. Although Defendant 2 presented a data authorisation letter to Defendant 3, the letter was not signed by the plaintiff personally. As none of the defendants had obtained the plaintiff’s informed consent, the court held that certain defendants had infringed the plaintiff’s voice-related rights and interests.
(III) Regulating, in accordance with law, acts that use AI to infringe natural persons’ privacy rights, such as “online doxxing” and “human flesh searches”
The protection of natural persons’ privacy rights has long been provided for under core laws such as the Civil Code and the Criminal Law, as well as more recently under the Personal Information Protection Law. The Opinions, however, specifically emphasise that where AI technology is used to track and analyse publicly available information relating to a specific natural person, such as telephone numbers, online accounts and social media information, even if the information itself is public, the conduct should still be found to constitute an infringement of privacy rights if the synthesised data intrudes upon the peaceful enjoyment of the person’s private life.
(IV) Determining, in accordance with law, civil liability for infringement of personal information rights and interests during AI training
With respect to the source and legality of training data, Article 19 of the Regulations on the Administration of Network Data Security emphasises that network data processors providing GenAI services should strengthen security management of training data and related processing activities. Departmental rules, including Article 7 of the Interim Measures for the Administration of Generative Artificial Intelligence Services, Article 11 of the Interim Measures for the Administration of AI Anthropomorphic Interactive Services and Article 14 of the Provisions on the Administration of Deep Synthesis of Internet Information Services, have also successively set out specific requirements concerning the legality of training data. Supporting national standards, such as GB/T 45652-2025 Cybersecurity Technology – Security Specification for Generative Artificial Intelligence Pre-training and Optimisation Training Data, also provide recommended guidance.
The Opinions make clear that although the use of already disclosed personal information for training will generally not be found to infringe personal information rights and interests, separate individual consent should be obtained if such use exceeds a reasonable scope or has a material impact on personal rights and interests. In determining what constitutes a reasonable scope, the training party should comprehensively consider factors such as the purpose of processing personal information and the necessity and appropriateness of the model’s functions, the type and sensitivity of the personal information involved and its potential impact on personal rights and interests, and the circumstances in which the individual disclosed the information and the scope of use that could reasonably be expected.
Note that if an individual has expressly objected to the processing of their personal data for AI training when making such data public, or subsequently objects after becoming aware it is being used for AI model training, the objection would make such data processing illegitimate.
[Case for Example] In the online tort liability dispute between Liao and a technology and culture company, the defendant company used a series of videos featuring the plaintiff, without obtaining the plaintiff’s authorisation or consent, to create face-swapping templates and uploaded them to the relevant software. The Beijing Internet Court held that although the defendant had not deliberately distorted the plaintiff’s image and had not infringed the plaintiff’s portrait right, in the process of using AI technology to generate video templates from the plaintiff’s images, the defendant collected, used and analysed the plaintiff’s facial features and processed the plaintiff’s personal information without consent, thereby infringing the plaintiff’s personal information rights and interests. The court ordered the defendant to issue a written apology to the plaintiff and compensate the plaintiff for moral damages and non-financial, non-physical harm.
(V) Prudently determining the tort liability of generative AI service providers
Possibly with a view to protecting AI products and platforms, the Opinions introduce a mechanism similar to the “safe harbour” principle to rationalise the allocation of liability for GenAI service providers. Where a rights holder notifies a service provider of infringing content generated by an AI product, and the service provider fails to promptly take necessary measures such as stopping the generation of the infringing content, blocking the relevant prompts that generate infringing content and restricting or suspending services for infringing users, the service provider should bear tort liability for the resulting damage in accordance with law. However, if the prompt entered by a network user is itself infringing content, and the user maliciously attempts to induce the AI product to generate infringing content and harm others’ interests, that network user should bear tort liability in accordance with law. The Opinions draw a reasonable boundary around the liability to be borne by AI platforms, so as to avoid excessive liability that could impede industry development. Assessing whether safe harbour measures are adequate should be based on an assessment of their timeliness and the current standard of technological capabilities.
[Case for Example] In the online tort liability dispute between Li and Baidu, the plaintiff, a lawyer, was defamed by Baidu’s auto-complete suggestions, which associated the plaintiff with derogatory terms such as “sentenced to imprisonment” and thereby lowered the plaintiff’s social reputation. Baidu’s remedial measures also failed to eliminate the impact in a timely manner, so the plaintiff requested that Baidu be held liable. Taking into account Baidu’s degree of fault, the duration of the infringement, the consequences of the damage and other factors, the court ultimately held that Baidu’s conduct at issue had indeed infringed the plaintiff’s reputation right and ordered Baidu to issue a written apology to the plaintiff, while dismissing the plaintiff’s other claims.
(VI) Applying the injunction system for infringement of personality rights in accordance with law
Because AI-generated content is inexpensive to produce, spreads rapidly and can be repeatedly generated with ease, claiming damages only after the infringement has occurred may be insufficient to prevent the proliferation of harm. The Opinions therefore further clarify the application of injunctions for infringement of personality rights in AI-related infringement scenarios: where a rights holder has evidence proving that the infringing act is ongoing or imminent, and that failure to stop it in a timely manner would cause irreparable harm, the rights holder may apply to the court for an order requiring the actor to cease the relevant conduct, or requiring the relevant platform to stop providing the relevant services.
When applying for an injunction, the applicant should specifically explain the infringing content, its dissemination and the urgency of the harm, and propose measures proportionate to the relevant risk. For example, the applicant may request that the dissemination of a specific infringing video be stopped or that a specific account be restricted from using relevant functions, rather than directly requesting the shutdown of an entire platform or the blocking of multiple keywords. At the same time, legal persons and unincorporated organisations may also apply for injunctions in accordance with law. Accordingly, where companies encounter AI-generated false statements or attacks on business reputation, they may also consider this remedy.
(VII) Accurately determining product liability for AI products
For AI products embodied in physical products, such as intelligent robots and smart home devices, the Opinions require their product attributes to be determined in accordance with the Product Quality Law and, where a product is defective and causes damage, the liability of the producer and seller should be determined in accordance with law. AI services that are not embodied in physical products do not fall within the AI product liability framework.
Unlike traditional products, AI products may change their mode of operation after autonomous learning or software updates. Therefore, the determination of defects involves not only the condition of the product at the time it leaves the factory, but also its intended use, autonomous learning capability, updates and the degree of user control. Whether the producer or seller has truthfully explained the applicable intended use scenarios and their inherent limitations, and has clearly warned of foreseeable risks, will also become a key focus of judicial review.
From a practical perspective, companies should continuously retain records of product testing, version updates and sufficiently clear risk disclosures. After an incident occurs, they should also promptly preserve the software version, operating logs and operational circumstances at the relevant time, so as to prevent subsequent updates from altering key evidence.
Note that while failure to comply with an applicable industry safety standard may be evidence of a defect in an AI product, even if an AI product complies with such standards, it could be found to be defective where the AI product itself poses an unreasonable danger to personal safety or property.
(VIII) Regulating, in accordance with law, acts that infringe consumers’ lawful rights and interests, such as “big-data-enabled price discrimination against existing customers” and “impersonating celebrities for livestream selling”
With respect to “big-data-enabled price discrimination against existing customers”, the Opinions do not treat all differential pricing as infringing conduct. Instead, they require an examination of whether the differential treatment has legitimate, sufficient and non-discriminatory grounds. Factors to be considered include whether the business operator forms individualised transaction terms based on consumers’ willingness to pay, browsing records and other information, and whether the relevant arrangements substantively impair consumers’ rights to know, to make independent choices and to conduct fair transactions. Therefore, the mere fact that two accounts display different prices will generally be insufficient to prove “big-data-enabled price discrimination against existing customers”. It remains necessary to exclude the effects of factors such as transaction timing, discount eligibility and service content.
Regarding the use of AI to impersonate celebrities for livestream selling, the Opinions further clarify that, where such conduct constitutes fraud, consumers may claim punitive damages under Article 55 of the Consumer Rights and Interests Protection Law.
[Case for Example] In the online tort liability dispute between Li and a cultural media company, the defendant used the plaintiff’s speech and teaching videos when promoting family education books and paired them with an AI-synthesised voice highly similar to the plaintiff’s voice, making it easy for the public to believe that the plaintiff, as a KOL, had participated in the book promotion. The court held that there was an entrusted promotion relationship between the business operator and the livestream seller, and that the business operator had the ability to review and manage the relevant promotion but failed to reasonably examine the authorisation status, and should therefore bear corresponding tort liability. Although the case concerned infringement of personality rights rather than a consumer claim for punitive damages, it also shows that business operators may bear corresponding liability for “impersonating celebrities for livestream selling”.
(IX) Determining, in accordance with law, liability for compensation in traffic accidents involving autonomous vehicles and vehicles with assisted-driving functions
For accidents involving autonomous driving and assisted driving, the Opinions continue to take the Civil Code and the Road Traffic Safety Law as the basis for determining liability. Where damage is caused by a vehicle defect, the injured party may request the producer or seller to bear product liability in accordance with law. Where the same damage involving a vehicle with assisted-driving functions is jointly caused by a product defect and the driver’s fault, liability should be allocated in accordance with law by considering the conduct of each party and its causal contribution. The driver’s liability should not be automatically exempted merely because the assisted-driving function was activated.
At the same time, accident liability and liability for sales promotion should be analysed separately. Where producers or sellers make false or misleading representations regarding the level of automation, performance or intended use, they may incur liability in relation to consumer rights protection even if the issue is not limited to the accident itself.
The difficulty in such cases usually lies in reconstructing the accident process. The Opinions make clear that courts may require the data controller, within the necessary scope, to provide authentic and complete data needed to ascertain the facts of the case, such as event records relating to autonomous driving or assisted driving. In our view, relevant stakeholders may also rely on this point when applying to the court for an investigation order to protect their rights and interests.
III. AI-related intellectual property cases
(I) Regulating, in accordance with law, acts involving AI-related infringement of intellectual property rights
Where AI-generated content infringes copyright, developers, providers and users should not automatically assume exactly the same liability merely because they are part of the same technological chain. The Opinions require liability to be reasonably determined by comprehensively considering the type of service, the source of training data, the degree of participation of each party, necessary measures taken and profits obtained. In particular, where a user knows or should have known of the existence of a prior work, but nevertheless uses AI to generate substantially similar content without any reasonable defence, the user should bear liability in accordance with law.
In terms of evidence, where a developer raises a non-infringement defence, the Opinions require the developer to provide materials such as the sources of training data, records of the training process, the model’s operating mode and the underlying scientific theories i.e. the burden of proof remains the same. It is therefore particularly important for AI product developers to conduct sufficient compliance work at the R&D stage and retain AI product-related materials.
[Case for Example] In the Hangzhou “Ultraman” case, the court considered the model and image-sharing functions provided by the platform, the infringement risks and the platform’s performance of its duty of care, and held that the platform had committed contributory infringement of the right of communication through information networks. The court ordered the platform to cease the infringement and pay RMB 30,000 in economic losses and reasonable expenses. The second-instance court upheld the original judgment. This case shows that whether a platform should bear liability depends on the functions it actually provides and its degree of participation in the infringing activities, and cannot be determined merely by reference to an assertion of “technological neutrality”.
Note that the Opinions do not address certain key IP issues, such as the copyrightability of AI-generated content and the issue of whether the use of copyrighted works for AI model training purposes without permission constitutes fair use.
(II) Reasonably determining, in accordance with law, the legal liability of parties involved in AI-related open-source software
Open-source software lowers the threshold for AI research and development, but the free provision of software and disclosure of source code do not necessarily exempt developers from all liability for subsequent infringements. The Opinions require appropriate liability exemptions to be granted to developers and providers by considering factors such as the type of open-source licence agreement, restrictions on rights, safety and compliance measures, and the degree of information disclosure. Where a party provides certain code modules free of charge and publicly explains their functions and safety risks, and another party’s use of such modules results in infringement, the court may determine that the original developer or provider does not bear liability.
Accordingly, developers of open-source projects should specify licences, use restrictions and risk disclosures in writing. Companies using open-source components should verify authorisation conditions, retain version records and review whether their own modification, integration and deployment activities increase risks.
(III) Regulating, in accordance with law, patent granting and validity confirmation for AI-related inventions and creations
The Opinions set out principled rules on the subject matter of patent protection for AI-related inventions and creations. The use of AI to assist R&D will not automatically exclude patent protection, but users should not be recognised as inventors merely because they operated AI tools. Companies should preserve records showing how R&D personnel identified problems, selected technical solutions, designed experiments and modified results, and should fulfil the sufficient disclosure requirement by fully explaining in application documents the technical content necessary to implement the solution. This helps avoid a situation where model outputs are directly compiled into a patent application, but the applicant is unable to explain the actual human contribution or the implementation of the technical solution.
(IV) Prudently determining, in accordance with law, liability for breach of contract in AI-related technology contracts
AI technology R&D involves a certain degree of uncertainty and model performance may also be affected by training data, application scenarios and evaluation methods. Therefore, when hearing disputes involving technology development, transfer, licensing, consulting and service contracts, the Opinions require courts to determine liability for breach of contract in accordance with the contractual arrangements, while fully considering the characteristics of R&D and whether the developer has made reasonable efforts.
From a contractual perspective, the parties should, as far as possible, clearly define the testing data, evaluation methods, acceptance criteria and applicable scenarios, and distinguish between deliverables that are promised to be delivered and R&D objectives that require continued exploration. The parties should also agree on cooperation obligations relating to data provision, changes in requirements, version upgrades and defect remediation. After a dispute arises, it is necessary to ascertain whether the failure to meet performance standards resulted from the developer’s failure to perform as agreed, or from the quality of customer data, changes in requirements or reasonable R&D risks, so as to avoid allocating all adverse consequences to one party.
(V) Regulating, in accordance with law, AI-related data use
The Opinions clarify the corresponding protection paths based on the different characteristics of data and data collections: where the requirements for constituting a work are met, the Copyright Law may apply; where the requirements for trade secret protection are met, the Anti-Unfair Competition Law may apply; and where the data does not constitute a trade secret but the alleged conduct violates Article 13 of the Anti-Unfair Competition Law, liability should also be borne in accordance with law. Therefore, data protection is not premised on registration, but companies still need to prove the lawful source of the data, the specific rights and interests involved, and the unlawfulness of the alleged conduct.
[Case for Example] In the unfair competition dispute concerning the “cartoon transformation special effect”, the defendant copied a cartoon special-effect model developed by the plaintiff with substantial operating resources and placed it on the market, competing with the plaintiff for traffic and users. The court supported protection of the competitive interests embodied in the structure and parameters of the AI model, and ordered the defendant to cease the infringement, eliminate the impact and compensate the plaintiff for economic losses and reasonable expenses totalling more than RMB 5 million.
Conclusion
Overall, the Opinions further clarify, within the existing legal framework, the direction for determining tort liability, intellectual property liability and contractual liability, based on different AI application scenarios and the degree of participation of each party. At the same time, Part IV of the Opinions sets out requirements for ascertaining technical facts, reviewing electronic evidence and verifying AI-assisted litigation materials, while Part V promotes the improvement of rules through diversified dispute resolution, case guidance and inter-departmental coordination. It is likely that future disputes will place greater emphasis on companies’ actual traceable records of actions taken. Companies are therefore advised to establish necessary preventive mechanisms with reference to the Opinions to avoid a lack of written evidential materials when disputes arise.
For further information, please contact :
Dominic Edmondson, Partner, Dentons
