China’s Supreme People’s Court has told judges how to decide AI disputes without waiting for an AI law. On 7 September 2026 the court published the Opinions on lawfully hearing AI-related dispute cases (最高人民法院关于依法审理涉人工智能纠纷案件的意见), document Fa Fa [2026] No. 10, in five parts and 24 articles. Vice President Tao Kaiyuan released them at a press conference in Beijing.
The timing matters because China still has no dedicated AI statute. The court says so in its own release, and builds the Opinions on the Civil Code, the Cybersecurity, Data Security, Copyright, Anti-Unfair Competition, Consumer Protection and Personal Information Protection laws, and the Civil Procedure Law. It calls the text the first set of AI judicial rules issued by the country’s top court. For anyone building or deploying generative AI in China, this is now the clearest statement of how courts will weigh fault, evidence and responsibility.
The Supreme People’s Court issued guidance, not an AI Act
Opinions of this kind from the Supreme People’s Court tell lower courts how to apply existing law; they do not create a new statute passed by the legislature. Article 2 sets three principles: put people first, support innovation and hold a safety baseline. It also asks courts to treat general-purpose and specialised models, and open and closed models, differently when sharing out responsibility between developers, providers and users.
China has been building AI policy on several tracks at once. TSN reported in October on a separate guideline telling officials to cushion AI job losses, which, like these Opinions, works through existing institutions rather than a single AI law.
Liability starts from fault unless a law says otherwise
Article 3 applies the ordinary fault rule in Civil Code article 1165(1) unless another law sets strict or presumed-fault liability. To judge fault, courts weigh the setting, the system’s degree of autonomy, transparency, the risks involved, the preventive steps a developer or provider took, and how far the user could foresee and control the harm.
Physical products are treated differently. Article 9 brings AI products with a physical carrier under the Product Quality Law, so producers and sellers bear product liability for defects. Courts are told to weigh the product’s purpose, its self-learning ability, updates, how much control the user has and compliance with standards, and to look closely at whether makers warned clearly about limits and foreseeable risks. Jones Day’s commentary describes this as strict liability; the official text speaks of product liability under existing law, so “strict” is the firm’s gloss.
Generative AI providers must act on notice
Articles 4 and 5 deal with the harms people most often meet. Using someone’s face or voice to generate an identifiable digital likeness without consent can infringe personality rights, and so can “AI resurrection” of the dead, which close relatives may challenge. Using AI to track and expose a person’s private information, the practice known in China as “opening the box” or human-flesh searching, infringes privacy. TSN has covered similar moves elsewhere, such as Queensland’s bill to make creating sexual deepfakes a crime.
Article 7 sets a notice-and-takedown duty. A generative AI provider that fails to act promptly after a valid notice, which must include preliminary evidence and the rights holder’s real identity, is liable for the resulting harm. Users who maliciously prompt a model into producing infringing output are liable too. Article 8 lets courts issue personality-rights injunctions against an AI provider where harm is imminent.
Article 6 covers personal data in training. Processing personal information that a person made public, or that was otherwise lawfully public, within a reasonable scope and without an express refusal, is generally not infringement. Consent is still required where there is a major impact on the individual.
Developers claiming non-infringement must show their training data
Article 12 is the provision most likely to be felt in copyright cases. When an AI developer argues it did not infringe, the court should order it to produce evidence of its training data sources, training records, how the model runs and the scientific basis. In the court’s question-and-answer briefing, officials said the fact that content was AI-generated is no reason to excuse anyone, and that liability should follow each party’s ability to control the output and its duty of care. A user who knowingly generates a work substantially similar to an earlier one is liable.
Article 13 offers a safe harbour for open-source developers that release free code modules and disclose their functions and security risks, if someone else uses the code to infringe. Article 14 says AI-assisted inventions can be patentable, with the natural person who made a creative contribution named as inventor. Article 16 protects lawfully obtained training data and datasets through copyright, trade secrets or unfair-competition law.
Courts can demand data and punish fake AI evidence
Article 11 covers crashes involving autonomous and assisted-driving cars. Courts may require makers, sellers or operators to supply driving event data, and carmakers that overstate automation levels can be liable to consumers. Article 10 targets algorithmic price discrimination and lets consumers seek punitive damages when AI fake-celebrity endorsements amount to fraud.
Articles 17 to 19 turn to procedure. Courts can draw an adverse inference if a party withholds evidence it controls, and must scrutinise AI-generated content offered as proof. Article 19 says litigants who fabricate evidence by manipulating AI can have claims dismissed and face fines or detention, and that anyone filing AI-generated documents or case-search reports must verify them, disclose the AI assistance and answer for their accuracy.
Why the training-data evidence rule matters more than the fault rule
The headline principle, fault-based liability, looks developer-friendly. In practice, the burden may sit elsewhere. Copyright and data claims against model makers usually fail or succeed on what went into the model, and claimants rarely have that information. By requiring a developer that pleads non-infringement to produce training sources and records, Article 12 shifts the practical burden of proof to the party that holds the evidence. The Supreme People’s Court’s own briefing frames this as a response to the complexity and opacity of AI systems.
That approach sits between the US and European paths. In US cases such as the dispute over whether Suno deleted its training recordings and the individual claims some 300 authors have brought against Meta, training data is pried out through discovery fights. The SPC has written the expectation into guidance from the outset. For companies operating in China, good records of data sources become a defence asset, not just a compliance cost.
The open-source safe harbour points the same way: disclosure earns protection. Two limits remain. These are Opinions, so they guide courts rather than bind every outcome, and the SPC says it deliberately left contested issues blank until more cases accumulate. The Opinions do not settle whether AI-generated content can itself be protected by copyright.
Jones Day’s English summary is useful but should be read with care. It omits Article 6 and Article 19, and presents a 2021 Shaoxing case as if it applied an article of this 2026 document; it is better read as an earlier example of the same kind of dispute.
The next test for AI disputes is in lower courts: the first judgments that cite Fa Fa [2026] No. 10 will show whether Article 12’s training-data duty is enforced as written.
Updated 10 November 2026: Added the document number and the SPC’s question-and-answer briefing from the court’s official pages, added Articles 8, 10, 16 and 17-19 from the official text, and replaced the summary and open-questions sections with analysis.
Sources
- Supreme People’s Court: release and full text of the Opinions on AI-related disputes, 7 September 2026 (Chinese)
- Supreme People’s Court: officials answer reporters’ questions on the Opinions (Chinese)
- Jones Day: China’s Supreme Court issues landmark judicial rules on artificial intelligence (September 2026)
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterFollow @tsnmediahub on X for breaking AI news as it happens.
Follow on X


