Authors
Background
On 7 September 2026, the Supreme People’s Court of China (SPC) published the Opinions on the Trial of AI-Related Dispute Cases According to Law (the Opinions). This is the first judicial rule-making document on AI disputes issued by a national supreme judicial body in China. It marks a significant step towards a structured approach to the novel legal questions arising from the rapid development and deployment of artificial intelligence.
China regulates AI through a multi-layered framework of broad foundational statutes and targeted administrative measures, rather than through a single omnibus national law. Against that backdrop, the SPC drew on a broad range of existing statutes to formulate the Opinions, including 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. The Opinions comprise 24 articles organised into five parts: (i) guiding principles; (ii) AI tort liability; (iii) AI intellectual property disputes; (iv) procedural and evidentiary rules and criminal liability; and (v) working mechanisms.
The Opinions respond to a rapidly evolving landscape in which new categories of disputes have emerged, including disputes involving deepfakes, AI hallucinations, big data price discrimination, autonomous vehicle accidents, and intellectual property issues arising from model training.
Key provisions of the judicial opinions
Determination of liability
Liability allocation and fault-based responsibility. The Opinions establish fault-based liability as the default regime for AI-related disputes. They identify six factors for determining fault: the application scenario; the degree of autonomy of the AI system; technical transparency; the potential risks and scope of impact; the preventive measures taken by the AI developer and service provider; and the user’s ability to foresee and control harm. The Opinions also distinguish between general-purpose and specialised models, and between open-source and closed-source systems, recognising that each category presents a different risk profile.
AI product liability. The Opinions define “AI products” narrowly as physical products, such as robots and autonomous vehicles. Pure AI services fall outside the product liability regime. Product defect assessments consider the product’s nature and purpose, autonomous learning capability, update status, degree of user control, compliance with applicable standards, and whether the manufacturer and seller provided adequate warnings.
Generative AI service provider liability. The Opinions apply a “safe harbour” notice-and-takedown rule by analogy to generative AI service providers. Providers are not liable if they take timely and necessary measures after receiving notification from a rights holder. Necessary measures include stopping the generation of infringing content and blocking related prompts. The so-called “red flag” rule, which would have imposed proactive monitoring obligations, was discussed during the drafting process but was ultimately excluded from the final text of the Opinions.
Autonomous vehicles. Existing traffic-accident liability rules apply to autonomous and assisted-driving vehicles. If an accident is caused by a vehicle defect, product liability applies. Where both a product defect and driver error contribute to the harm, the manufacturer and driver may be held jointly liable. Manufacturers, sellers, operators, and other data controllers must provide vehicle data to courts to facilitate fact-finding.
Personal information and privacy protection
Personality rights protection. The Opinions address AI-enabled infringements of personality rights, including deepfakes such as face-swapping and voice cloning. Using AI without consent to generate an identifiable virtual image of a person may infringe portrait, name, or voice rights. The Opinions also protect the personality interests of deceased persons against the so-called “AI resurrection” of the deceased. They reinforce the concept of “cyber doxxing”: using AI to track and analyse publicly available information to obtain an individual’s private information constitutes a privacy infringement. In urgent cases, the personality-rights injunction system provides interim relief.
Personal information and training data. Processing publicly available personal information for model training generally falls within a “reasonable scope” and does not constitute infringement, provided the individual has not explicitly objected. The assessment considers the purpose and necessity of the processing, the sensitivity of the information, and the data subject’s reasonable expectations. However, informed consent is required where the processing has a “significant impact” on individual rights.
IP and consumer protection
Intellectual property. Where AI-generated content infringes copyright, responsibility is allocated among developers, service providers, and users based on fault. Relevant considerations include the type of service, training-data sources, each party’s level of participation, preventive measures taken, and profits earned. Open-source software developers and providers may benefit from liability exemptions if they provide code free of charge and publicly disclose the functionality and security risks of their models. The Opinions confirm that AI-assisted inventions may be patented, provided that a natural person makes a substantive creative contribution. Two major issues remain deliberately unresolved: whether AI-generated content is eligible for copyright protection, and whether using copyrighted works to train models constitutes infringement. For AI-related breach-of-contract claims, liability should be assessed by considering the characteristics of the technology and whether the developer exercised reasonable efforts.
Consumer protection. The Opinions address “big data price discrimination”, in which user profiling produces differential pricing that materially harms consumers’ rights, including their rights to information, choice, and fair dealing. Such conduct constitutes unreasonable differential treatment under the law. A service provider bears corresponding tort liability where the price discrimination infringes consumers’ lawful rights and interests.
Evidentiary issues in the AI era
Courts must account for the complexity and opacity of AI systems when evaluating evidence. They assess AI-generated evidence by reference to prompt design, the similarity of outputs to the claimed works, the repeatability of tests, and the model’s training and filtering mechanisms. Parties that submit AI-generated materials to a court must verify their accuracy and disclose the use of AI. Fabricating evidence with AI may result in fines, detention, or criminal prosecution.
Business implications for companies
The Opinions mark a shift from static compliance to “provable compliance” across the full AI lifecycle. By turning technical governance expectations into judicially enforceable duties of care and evidentiary obligations, they require businesses to document, validate, and defend their AI-related practices.
For organisations using open-source models, the Opinions require review mechanisms covering licence terms, commercial-use restrictions, acceptable-use policies, and disclosed security risks.
AI development, licensing, and procurement agreements should include service-level requirements, hallucination-rate benchmarks, safety guardrails, and human-oversight mechanisms.
For training-data governance, enterprises must conduct privacy impact assessments for model training using publicly available data. They should also maintain complete audit trails covering data sources, cleansing rules, and authorisation documents. These records may become litigation evidence.
For physical AI products, safety assessments must continue throughout the product lifecycle and cover software updates and model versioning. Litigation readiness is equally important. Enterprises should preserve comprehensive records, including prompts, system prompts, retrieval-augmented-generation materials, generation logs, and human-review records, because courts may require them as evidence.
Users who knowingly prompt AI to generate content substantially similar to existing works may face infringement liability. Enterprises should therefore implement appropriate content-review workflows.
Key practical steps
Given that the Opinions will serve as the authoritative basis for judicial decisions in AI-related disputes, companies operating in China should treat compliance as an immediate priority. Companies should promptly align their internal policies, contracts, and governance frameworks with the Opinions:
- Assess gaps and monitor the regulatory landscape. Companies must map all AI applications used in services and products against the Opinions to identify compliance gaps across fault-based, presumed-fault, and strict-liability regimes. China’s draft AI regulations, TC260 technical standards, and subsequent judicial interpretations may further supplement the framework.
- Build content monitoring and response workflows. Companies should establish a process for receiving and acting on rights-holder notifications, including the ability to stop infringing content generation and block related prompts. It is also good practice to add pre-publication review, similarity checks against existing works, and brand-compliance checks.
- Enhance technical and safety governance. Businesses should conduct ongoing defect and safety assessments for physical AI products such as robots and autonomous vehicles, covering software updates, model versioning, and user control throughout the product lifecycle. Training-data governance calls for privacy impact assessments, documented data sources, and complete audit trails.
- Strengthen contracts and litigation readiness. AI procurement, development, and licensing agreements should include model performance standards, safety guardrails, data-provenance obligations, and evidence-preservation requirements. Generation records, including prompts, model versions, timestamps, and human-review logs, should be preserved for potential court proceedings.
- Manage rights and consent. Businesses should consider obtaining specific, documented consent before using individuals’ voices, likenesses, or other personal attributes in AI training or content generation. For publicly available personal information, the processing must fall within a reasonable scope, and informed consent in compliance with relevant regulations is required where it has a significant impact on individual rights.
Client Alert 2026-192