On September 7, the Supreme Court wrote the rules down
China's Supreme People's Court issued its Opinion on Lawfully Adjudicating Disputes Involving Artificial Intelligence: five parts, 24 articles, the first AI adjudication rules from the country's highest court. China still has no dedicated AI law, so courts facing a wave of novel cases have had no common yardstick. This fills the gap for now.
It covers AI face swapping and voice cloning, resurrection of the dead, harm from hallucinations, doxxing, algorithmic price discrimination, celebrity impersonation in commerce, accidents involving autonomous and assisted driving, use of public personal information in training, open-source software, product liability, and AI-generated evidence. Below are the parts that matter most to builders.
Face swap and voice cloning: allowed, but get consent first
Using generative AI to process the name or likeness of a person, or of a deceased person, must not break the law or offend public order and good morals. Unless law provides otherwise, generating an identifiable digital likeness or synthesized voice without that person's consent supports an infringement claim in court. Using a digital human for improper conduct or false statements that harm someone is also infringement. Shen Weixing of Tsinghua put it simply: you may use it, but you need authorization.
The personality rights injunction is now usable
Article 8 drew the most attention. With evidence that someone is using or about to use AI to infringe personality rights, and that delay would cause irreparable harm, a party can ask a court to order the conduct stopped, or to order the network provider or the generative AI provider to stop serving. The court's reasoning was practical: AI harm is cheap, fast, wide and hard to reverse, so waiting for a full trial is often too late. The example given was a victim of fabricated explicit deepfake content.
There is a written counterweight. Injunctions must fit the right at stake, the manner of violation and the likely harm, and may not exceed what is necessary.
Provider liability: the safe harbor arrives
For generative AI providers, the opinion imports the Civil Code's notice-and-takedown rule. A provider cannot foresee everything users type, and output volume makes item-by-item review impossible. But once a valid notice arrives it must act promptly, or it shares liability for the expanded harm. Act promptly and it generally owes nothing.
For teams shipping AI products this is a direct instruction: build a working complaint channel and a fast takedown path. Without both, the safe harbor does not cover you.
Product liability and self-driving
Article 9: AI products in physical form are assessed under the Product Quality Law. Where a defect causes harm, producers and sellers bear liability. Courts weigh the product's nature and use, its autonomous learning, its update history, user control and standards compliance, and specifically examine whether the maker gave truthful descriptions and clear warnings about applicable scenarios, inherent limitations and foreseeable risks. Whether your docs say where the product fails will shape how liability is assigned.
Article 11: accidents involving autonomous or assisted vehicles follow the Civil Code and Road Traffic Safety Law; where a defect caused the crash, claims go to the producer or seller. Shen reads assisted driving as shared human-and-machine responsibility.
The opinion also covers algorithmic price discrimination, celebrity impersonation, where consumers may claim punitive damages under Article 55 of the Consumer Rights Protection Law if fraud is established, and doxxing through AI-assisted analysis of public information, which counts as a privacy violation.
Two big questions left blank
Issues without consensus are left open until experience accumulates. Two of them: whether AI-generated content has copyright, and how to treat unlicensed use of copyrighted works in training. These are exactly what industry wants answered, the second especially, since it defines the compliance line for training data. Blank does not mean ignored. It means courts may diverge for a while.
What this means for you
Three checks worth doing now. One: for anything touching faces, voices or likenesses, complete the authorization chain and keep records. Digital humans, voice cloning and face swap are all in range. Two: build the complaint channel and fast takedown path, your precondition for the safe harbor and the cheapest protection available. Three: rewrite the docs. Scenarios, limitations, foreseeable risks.
On the two open questions, do not bet on a direction. Keep evidence and records, so when rules arrive you can show what you did.