On September 7 2026, the Supreme People’s Court of China released the Opinions on the Adjudication of AI-Related Disputes in Accordance with Law (the “Opinions”). Although the Opinions do not constitute a standalone AI statute, they are the most comprehensive judicial guidance issued to date on how Chinese courts should apply existing civil (tort), intellectual property (IP), consumer protection, anti-unfair competition, data protection and procedural rules to AI-related disputes. The Opinions cover a broad spectrum of issues, including AI-generated impersonation, hallucinated outputs, doxxing, algorithmic discrimination, autonomous driving, AI training data, copyright and patent disputes, open-source AI software, data rights and evidentiary review. The Opinions seek to provide clear guidance on several emerging and highly debated legal issues involving AI, including whether safe harbor rules may apply by analogy to generative AI service providers and the evidentiary rules applicable to copyright infringement disputes involving AI-generated content. At the same time, however, the Opinions remain silent for the time being on issues that continue to be subject to significant debate, such as the copyrightability of AI-generated content and the legal characterization of the unauthorized use of copyrighted works to train large AI models.

Companies developing or deploying AI in China should therefore treat the guidance as a litigation-risk roadmap: AI governance and algorithm supervision records, transparent disclosures and consent chains, training-data compliance provenance, product warnings, notice-and-consent mechanism, platform controls and evidentiary logs are likely to become central to both liability allocation and dispute defense.

1. Primary legal framework remains unchanged

The Opinions confirm that Chinese courts will adjudicate AI-related disputes primarily by applying existing laws, including the Civil Code, Cybersecurity Law, Data Security Law, Copyright Law, Patent Law, Anti-Unfair Competition Law, Road Traffic Safety Law, Consumer Protection Law, Personal Information Protection Law and Civil Procedure Law. The Supreme People’s Court noted that they may release further guidance which signals a deliberate phased approach: while the existing clear rules demand immediate compliance, the unresolved areas carry litigation uncertainty and also present opportunities for businesses to help shape future jurisprudence through active case participation and continuous monitoring of subsequent judicial interpretations, guiding cases, or legislative developments.

2. Key substantive rules and risk areas

(1) Tortious liability and the importance of risk assessment documents as evidence

Plaintiffs generally bear the burden of proving fault in tort claims. In AI-related cases, however, the Opinions advise Chinese courts to assess fault by considering AI-specific factors, including the application scenario, degree of autonomy, technological and informational transparency, potential risks and impact, preventive measures adopted by developers and providers, technical feasibility of risk mitigation, and the user’s ability to foresee and control harm. Companies should therefore maintain pre-launch AI risk assessments, data protection impact assessments, model governance records, model update logs, human oversight arrangements and incident-response documentation, as these materials may be critical evidence in defending against negligence claims. Given the rapid development of AI technologies and business operations, and the possibility that such evidence may include internal documents or trade secrets, companies should also prepare strategies for evidence preservation and court disclosure.

(2) Deepfake, voice cloning and personality rights

Consent remains the central compliance requirement where AI is used to process a person’s name, portrait, voice or other identity-related features. Courts may find infringement where AI is used without authorization or consent to create or publish identifiable virtual images, clone a person’s voice, manipulate a digital image or voice to make false statements or improper conduct, or create unauthorized digital representations of deceased persons. Collecting voice or image datasets of celebrities or deceased persons without consent may therefore raise legal concerns, particularly where the data could be used for voice cloning or image and voice manipulation. If the relevant technology or algorithm is used in telephone or online scams, law enforcement may examine the purpose of the data collection and product development, and criminal liability may arise where the technology is knowingly supplied to such criminal actors. Companies using avatar, voice synthesis, celebrity-like marketing or digital human technologies should review consent chains, training data provenance, contractual warranties and takedown procedures.

(3) AI training data and publicly available personal information

The Opinions provide a relatively practical position on model training: processing personal information that an individual has voluntarily disclosed, or that has otherwise been lawfully disclosed, will generally not be treated as infringing personal information rights if it is used within a reasonable scope and the individual has not expressly objected or opted out. However, consent may still be required where the processing has a significant impact on individual rights. The “reasonable scope” analysis will consider necessity and proportionality, the type and sensitivity of the data, potential impact on individuals, and the context and reasonable expectation when the information was made public. This is likely to increase the importance of opt-out mechanisms, data impact assessment and documented training-data governance.

(4) Platform liability for AI providers

The Opinions take a calibrated approach to platform liability. Where a user uses AI-generated content to mislead others or infringe rights, the user may bear primary responsibility. A platform’s liability is more likely to arise where it receives valid notice of the infringement, and fails to take necessary measures such as takedown, blocking, disconnection or other reasonable controls. Platforms should therefore strengthen notice-and-action mechanisms, user terms, evidence retention and repeat-infringer handling, rather than assuming strict liability for all user-generated AI content.

(5) Product liability

AI products that contain defects and cause damage may trigger product liability for producers and sellers. In assessing whether an AI product presents an unreasonable danger to personal or property safety, courts shall also take into consideration AI-related features, such as the specific nature and use of the product, autonomous learning capacity, updates and upgrades, user’s ability to control, applicable national or industry standards, and, most importantly as pointed out by the Opinions, whether the producer or seller has provided truthful explanations and clear warnings on intended use, inherent limitations and foreseeable risks. Although the Opinions do not yet resolve all questions on what constitutes a “defect” in AI products, they point companies toward clearer labeling, user instructions, risk warnings, testing records and post-market monitoring. As in PRC judicial practice, a “warning defect” is normally much easier to be established by the plaintiffs comparing with a “manufacture defect” and a “design defect” due to the lower technical barrier, while the Opinions specifically emphasizes the importance of proper warning in AI products, the company shall carefully review and ensure the validity of their warning mechanism (e.g., whether the warning terms are up to date, comprehensive,  understandable, and does not fall into the scope of invalid standard terms or liability limitation clause).

(6) Consumer protection and liability arising from misrepresentation and lack of disclosure

Operators that use AI for misleading celebrity endorsements, synthetic-person marketing or similar fraudulent sales practices may face consumer claims for punitive damages under the Consumer Protection Law. The Opinions also address algorithmic discriminatory pricing: unreasonable differential treatment using algorithms may give rise to tort liability where it substantially restricts or harms consumers’ rights to know, choose independently or trade fairly, particularly where transaction terms are generated based on consumption preferences, willingness or ability to pay, browsing history or similar personal profiles.

(7) Liability arising from autonomous driving

For autonomous and assisted driving, the Opinions confirm that liability for traffic accidents will be allocated under the Civil Code and Road Traffic Safety Law, while product defect claims against producers or sellers may proceed where vehicle defects caused the accident. Importantly, courts may require producers, sellers, operators or other data controllers to provide authentic and complete autonomous-driving or assisted-driving event records within the necessary scope to determine accident causation. This may reduce evidentiary barriers for claimants and reinforce the need for OEMs and mobility operators to maintain reliable logging, data integrity, incident reconstruction and litigation-readiness protocols, and establish practical data provision protocol for such evidentiary data request by judicial bodies.

(8) IP infringement

With respect to IP infringement cases involving AI-generated content, the Opinions first identify the factors that courts should consider in determining copyright infringement. Specifically, the factors to be considered by courts include the service type, industry characteristics, training-data sources, degree of participation by relevant parties, preventive measures adopted and profit derived.

Second, the Opinions further clarify the allocation of the burden of proof in such cases. While rights holders are required to produce evidence of infringement, AI developers are expected to substantiate their non-infringement defenses with evidence concerning, among other things, the sources of training data, records of the training process, the model’s operating mechanisms, and the underlying scientific principles. Users may also face liability where they know or should know of a prior work and use AI to generate substantially similar content without a reasonable defense. With respect to the AI generation process, theOpinionsset out rules only for determining whether the generated output constitutes infringement. They deliberately refrain from addressing the more controversial issue of whether the use of rights holders’ works as inputs for model training may itself constitute infringement.

Separately, AI-enabled counterfeiting, false advertising, fake traffic or order manipulation may trigger unfair competition liability.

(9) Open-source AI software

The Opinions recognize that developers or providers of open-source AI software may receive appropriate liability exemptions depending on the type of open-source license, the scope of rights restrictions, the security and compliance measures adopted, and the level of disclosure provided. Where a developer provides AI-related code modules free of charge on an open-source basis and publicly explains their functions and security risks, courts may find that the open-source developer is not liable for infringement caused by another party’s downstream use. TheOpinionsclarify the liability of open-source software developers and providers, reflecting an encouraging approach toward the open-source community under China’s legal framework. This is consistent with the SPC’s stated policy of avoiding imposing excessive liability at an early stage of AI technological and industrial development, which could otherwise discourage innovation. Companies should nevertheless continue to review open-source license obligations, model cards, security notices, provenance records and downstream use restrictions.

(10) Patentability of AI generated outpoints

Consistent with the current Patent Examination Guidelines issued by the China National Intellectual Property Administration (CNIPA), the Opinions provide that AI-related inventions may be patentable if they use technical means that comply with the laws of nature, solve technical problems and produce technical effects consistent with the laws of nature, provided that they do not violate any law, public morality or public interest, and that a natural person has made a substantive contribution to the invention. Under this general principle, the Patent Examination Guidelines set out detailed rules and examples for determining whether AI-related inventions constitute patent-eligible subject matter. Further guidance on the application of these rules to specific circumstances is expected to emerge through future judicial practice.

Second, a natural person using AI may be recognized as the inventor if that person made creative contributions to the substantive features of the invention. However, further judicial guidance will be needed to clarify whether, and to what extent, the use of AI in the inventive process must be disclosed, as well as how inventive contributions should be assessed.

Finally, with respect to the sufficiency of disclosure for AI-related patents, the Opinions confirm that the generally applicable legal standard continues to apply: the specification must disclose the invention in sufficient detail to enable a person skilled in the art to carry out the invention. Given the particular characteristics of AI-related technologies, however, we expect the application of this general standard in practice to present challenges and potentially give rise to disputes.

(11) Use of data

The Opinions also confirm protection for data and datasets lawfully obtained through collection, generation, derivation, transfer or licensing. Depending on their characteristics, data assets may be protected as copyrighted compilations, trade secrets or through anti-unfair competition rules. Malicious conduct such as fabricated interference data, malicious data labeling or adversarial sample attacks that harm AI operational security may also attract liability. Businesses should therefore align AI data acquisition, labeling, licensing and security controls with both data-rights protection and anti-abuse strategies.

3.  Guidance on evidence and litigation readiness

The Opinions empower courts to actively guide evidence production, order investigations on their own motion, and draw adverse inferences against parties that unjustifiably withhold relevant electronic data or technical records. Thus, maintaining comprehensive, auditable documentation of AI training data, model logs, and system operations is critical, not only to avoid adverse presumptions, but also to leverage expert assistants and technical examiners to explain complex AI mechanisms and strengthen your case.

The Opinions also provide specific evidentiary guidance. For big-data analysis reports, courts should examine the source of raw data, data-cleaning rules and scientific reliability of the analytical method. For blockchain evidence, courts should examine the authenticity of data before being recorded on-chain and the reliability of the technical platform. Where AI-generated content is used as infringement evidence, courts may consider prompt design, the influence of prompts on outputs, similarity between generated content and asserted works, consistency of repeated testing, and model training, algorithm design and content-filtering mechanisms. Companies should preserve prompts, outputs, model versions, logs, testing records and risk-control records in a way that can be explained in litigation.

4. Practical implications for companies

The Opinions are likely to become an important reference point for AI developers, deployers, platforms, OEMs, consumer-facing businesses and companies commercializing AI-generated content or data assets in China. Immediate priorities include mapping AI use cases and liability theories; reviewing consent and disclosure language for digital twin, voice and image synthesis; strengthening training-data provenance, opt-out and sensitive-data controls; updating platform notice-and-action procedures; enhancing product warnings and post-market monitoring; preserving AI system logs and litigation evidence; and revisiting contracts with AI vendors, customers, data suppliers and open-source contributors to allocate responsibility for data, IP, safety, security and regulatory compliance.