CHANG TSI
Insights
On 7 September 2026, the Supreme People's Court of the PRC released the Opinions on Adjudicating AI-Related Disputes in Accordance with Law, comprising 24 articles in five parts. It is the first set of AI adjudication rules issued by the SPC, and it gives courts at all levels a systematic text to work from.
China has no dedicated AI statute, and there is no consensus yet on whether one is necessary. Disputes, meanwhile, have kept arriving. By our count, more than sixty AI-related cases had been publicly reported in China as of August 2026, spanning copyrightability, platform liability, personality rights and data rights, and extending into criminal matters. Courts have been applying general civil, copyright and personal information law case by case, with inevitable variation in outcome.
The Opinions create no new category of rights and no new form of liability. They return to the Civil Code, the Cybersecurity Law, the Personal Information Protection Law and the Copyright Law to locate the basis for claims and for liability. The stated drafting principle is to make full use of existing legal provisions.
Where those provisions do not reach the new facts, the Opinions adapt them rather than replace them. Article 4, for example, sets identifiability as the threshold for AI use of a person's name, likeness or voice. Article 7 redefines the "necessary measures" owed under Article 1195 of the Civil Code: in a generative setting there is no fixed item of content to delete, so the obligation becomes stopping generation and blocking the relevant prompts. Article 12 allocates the burden of proof on training data and algorithms.
Fault liability is the default. Under Article 3, unless a statute specifically provides for strict or presumed liability, ordinary fault liability applies. Fault is assessed against the deployment scenario, the degree of autonomy, technical and informational transparency, the scope of foreseeable risk, and the measures the developer or provider actually took, as well as the user's capacity to foresee and control the harm. Two reported cases illustrate the point: a provider was found not liable where it had given clear warnings about the model's limitations, while a search platform was held liable for defamation where it had combined false and defamatory generated text with the claimant's photograph in a way that clearly identified him. The difference lay in how far each had involved itself in producing the output.
"AI product" is defined narrowly. Article 9 limits product liability to AI carried in a physical form. Robots and autonomous vehicles fall within it; AI services without a physical carrier do not. For a company selling both hardware and model services, the liability path differs by product form.
Liability is matched to control. Article 12 requires courts to allocate responsibility among developers, providers and users by reference to the service type, the source of training data, each party's degree of participation, the measures taken and the profits derived. The starting point is not who a party is, but what it actually did.
The evidential burden has shifted. A developer asserting non-infringement must produce the source of its training data, records of the training process, the model's operating mode and supporting technical reasoning. Read with the adverse-inference rule in Article 17, raising the defence may itself trigger disclosure. Companies that treat training data as a core trade secret should consider this before a dispute arises.
Human contribution affects patentability. Article 14 places the absence of substantive human contribution among the grounds for excluding subject matter, alongside violations of law and public morality. Where the human creative contribution occurred needs to be capable of explanation, which gives R&D records and iteration logs a bearing on the right itself.
The Opinions do not address the copyrightability of AI-generated content, or how the unauthorised use of copyrighted works for model training should be characterised. The SPC has explained that views diverged significantly during drafting and that understanding needs to develop further.
Both questions go to the existence and scope of rights rather than to the method of adjudication, and both are better settled by legislation. Chinese courts continue to decide them case by case. On copyrightability, outcomes have gone both ways, with negative findings generally resting on the limited intellectual input involved rather than on any denial that such content can be protected.
Whatever the technology, liability attaches to conduct that can be established on the evidence. The centre of gravity for compliance moves accordingly: from meeting the formal requirements of regulators, to maintaining a process capable of withstanding judicial scrutiny.
Risk disclosures, content filtering, complaint handling, records of training data and model operation, prompt and iteration logs, and clear acceptance criteria in development contracts are all material a party may need to rely on in court.
This note highlights only the principal points. For the full analysis, or to discuss the implications for your business in China, please contact us.
