China's Highest Court Set the AI Liability Checklist Lawmakers Have Not Written

China's Supreme People's Court has issued the country's first national set of adjudication rules for artificial intelligence disputes, a 24-article document that tells judges how to apply existing law to model training, generated content, synthetic identities and AI evidence.

The Opinions of the Supreme People's Court on Lawfully Adjudicating Cases Involving Artificial Intelligence Disputes, released on 7 September as Fa Fa [2026] No. 10, set rules for adjudicating AI disputes under existing law.

That distinction limits what the document itself does. The Supreme People's Court release describes it as an adjudication-rules document issued to guide courts in deciding AI disputes under existing law. It does not enact a standalone statute and it is not a formal Fa Shi judicial interpretation. Article 22 directs higher courts to take up novel or rule-setting AI cases and supervise lower courts, while judgments still rest on the underlying statutes.

The result is doctrine before legislation. Beijing has not created one general AI cause of action. Its highest court has instead turned broad duties already scattered across civil, privacy, copyright, consumer and procedural law into tests that can change conduct now.

Public data is not a blank cheque

Article 6 gives model developers a qualified safe path for personal information that an individual disclosed or that was otherwise lawfully made public. Processing that information for model training is generally not an infringement when it remains within a reasonable scope and the person has not expressly refused.

The qualification carries the rule. According to the court's text, judges must weigh the processing purpose, the model's function, the type and sensitivity of the information, its potential effect, and the context and expected use when it was disclosed. Consent is required when processing has a major effect on personal rights and interests.

That is not a licence to scrape anything reachable without a login. A provider needs a defensible record of how data became public, what a person could reasonably expect, whether they objected, and why the use fits the model. The operational burden moves upstream, into dataset provenance and exclusion controls.

The court stopped short of deciding the larger copyright fight. In its press conference response, it said disagreement remained too deep to rule on whether unlicensed training on protected works is infringement or whether AI-generated output can receive copyright. The personal-information test is concrete. The copyright status of training remains open.

A complaint now has to reach the generation layer

Article 7 adapts the Civil Code's notice-and-action rule to generative AI. When automatically generated content infringes reputation, privacy or another personality right, a provider that receives a valid notice and fails to act promptly can be liable for the resulting harm.

The notice must contain preliminary evidence of infringement and the rights holder's real identity. The expected response is also specific. The provider may need to stop the infringing content from being generated and block the relevant generation instructions, not merely remove one public post after it appears.

This changes the product requirement. A complaint system built for URLs and uploaded files is not enough when a prompt can recreate the harm. Providers need notices linked to prompt patterns, outputs and model versions. They also need logs showing what was blocked, when and by which technically feasible measure.

The user remains liable when malicious prompts induce harmful output. Under Article 7's notice route, exposure follows a substantiated notice and a failure to take prompt, technically feasible action. That route sits inside the Opinion's general fault framework, which asks what developers, providers and users could foresee and control rather than treating every AI service as strictly liable.

The practical unit of moderation is no longer only the published item. It is the ability to generate the item again.

Synthetic identity stays inside existing rights

Article 4 does not create a new property right in a face or voice. It applies existing rights to name, likeness, voice and reputation to synthetic output.

Using AI without consent to create, use or publish an identifiable virtual likeness can infringe name and likeness rights. Using a person's voice as training material to imitate recognisable timbre, intonation and pronunciation can infringe voice rights. Manipulating an identifiable likeness or cloned voice to perform improper acts or make false statements can become a reputation tort. Close relatives can also sue over unauthorised digital recreations that injure a deceased person's protected personality interests under Civil Code Article 994.

The line is identifiability plus legally relevant use, not ownership of every visual resemblance or vocal characteristic. A generic style control and a person-specific clone present different exposure. The claim still has to fit an established personality right.

Article 5 applies the same existing-rights method to privacy. AI analysis of public phone numbers, accounts or social-media data becomes a privacy infringement when it is used to extract or disclose private information, or to disturb private life. Public inputs can still produce a private-rights violation when aggregation changes what can be learned.

Courtroom controls may bite first

The procedural provisions are less conspicuous than the deepfake rules and may prove easier to enforce. Article 19 requires anyone submitting AI-generated pleadings or case-search reports to verify the cited law, interpretations and cases, disclose the use of AI, and accept responsibility for accuracy. Fabricated evidence can trigger dismissal, fines, detention or criminal referral depending on the conduct.

Articles 17 and 18 give judges a method for examining opaque systems. Courts may preserve technical evidence, draw adverse inferences when a party withholds electronic records, and test AI output by examining prompts, similarity, repeatability, training methods, algorithms and filtering. Article 12 separately says a developer relying on a copyright non-infringement defence should produce training-data sources, training records and information about how the model operates.

These are recordkeeping rules in practical form. A company that cannot reproduce an output, explain a filter or identify a training source enters court at a disadvantage.

China still lacks a single AI statute, and this Opinion deliberately leaves major questions unresolved. Its immediate force lies elsewhere. It gives courts a shared vocabulary for fault, reasonable scope, effective notice, identifiability and proof. Providers now know which operational facts judges will ask for. Rights holders know what a sufficient complaint must contain. Litigants know that an AI-generated citation is their evidence, not the model's problem.

Legislation may eventually consolidate those rules. Until then, Chinese AI liability will be written in the records companies keep, the notices they act on and the cases higher courts choose to pull upward.