On September 7, 2026, the Supreme People’s Court (SPC) issued the Opinions on Adjudicating Cases Involving Artificial Intelligence Disputes in Accordance with the Law (Fa Fa [2026] No. 10). The 24-article document took effect immediately. It is the SPC’s first dedicated guidance on how courts should decide AI disputes, and it applies to every case in a Chinese court, whoever the parties are.

It is not a regulatory compliance regime. Algorithm filing, content labeling and similar administrative duties sit elsewhere. What the Opinion does is tell judges how to apportion blame, who must produce what evidence, and where the existing statutes stop and judicial discretion begins. For companies selling AI products in China, training on Chinese data, or deploying generative AI locally, that is the layer that decides who wins a lawsuit.

1. Fault is the default, with one big exception

Article 3 sets the baseline. Unless a statute expressly provides for strict liability or presumed fault, liability for harm caused with AI is decided under the fault principle of Article 1165(1) of the Civil Code. Courts weigh the application scenario, the degree of autonomy, technical and informational transparency, the potential risk and its scope, the precautions the developer or provider took and could technically take, and the user’s ability to foresee and control the harm.

The exception matters for hardware makers. Under Article 9, the Opinion applies the Product Quality Law only to AI products with a physical carrier, such as robots and vehicles, and keeps generative AI services out of product liability. Article 11 confirms that autonomous and driver-assistance vehicles stay within product liability, and a manufacturer or seller that overstates a vehicle’s automation level can also be held liable. So the accurate summary is that fault governs AI software and services, while physical AI products remain under strict product liability.

For software companies, the practical point is that precautions are evidence. What you warned users about, what safeguards you built and what you did after incidents will all be weighed.

2. Notice-and-takedown for generative AI, but only for personality rights

Article 7 extends a notice-and-takedown structure to generative AI providers. It covers content that infringes reputation, privacy and similar personality rights. It does not cover copyright, which Article 12 handles differently (see section 3).

Under Article 7, a provider that fails to act promptly after notice, by stopping the generation of the infringing content, is liable for the resulting harm. The notice must contain preliminary evidence of infringement and the right holder’s real identity. A user who deliberately induces infringing output with crafted prompts is liable too. If the provider ignores a notice, the right holder can claim against both under Civil Code Article 1195.

Han Kun reports that the drafters’ explanatory article says two adjustments follow from how generative AI works. A notice can rely on simulated prompts or technical testing, since private one-to-one output has no public link, and the required response is forward-looking, meaning blocking the offending generation and related prompts rather than deleting a stored item. The text itself points the same way. Article 18 tells courts to weigh prompt design and the consistency of repeated tests when AI output is offered as evidence.

The Opinion does not adopt a “red flag” rule. Han Kun reports that the draft had a “knew or should have known” standard that was dropped after industry objections, and that the drafters stress the omission does not bar liability where a platform edits, curates or recommends infringing output.

3. Copyright: who has to prove what

Article 12 does not offer a safe harbor. It ties liability to control and duty of care, looking at the type of service, industry practice, the source of training data, each party’s role, the measures taken and the profits earned.

Its most consequential feature is the allocation of proof. A right holder must show the output was generated by the AI and is substantially similar to its work. A developer that pleads non-infringement can be ordered to produce its training data sources, training process records, how the model operates and the scientific rationale for its position. A user who knew or should have known of the earlier work, and generates something substantially similar without a reasonable defense, can be held liable.

Article 17 adds teeth. If a party controlling documents or electronic data refuses to produce them without justification, the court may accept the other side’s claim about what the evidence says. A developer that cannot trace its training data therefore faces a risk beyond a weak defense.

4. Training on personal information

This is easy to overlook, and it matters for anyone building models on Chinese data. Under Article 6, processing personal information that individuals made public themselves, or that is otherwise lawfully public, for model training within a reasonable scope is generally not an infringement, provided the person has not expressly refused. Where the processing significantly affects the individual’s interests, consent is required. Courts judge reasonableness by the purpose and necessity of the training, the type and sensitivity of the information, and the context in which it was published and the use people could reasonably expect.

5. Open source, patents and data

Open source. Article 13 directs courts to weigh the license type, its restrictions, safety and compliance measures and the degree of disclosure, and to grant open-source developers and providers an appropriate degree of exemption. A party that releases some code modules free and open source, and publicly explains their function and safety risks, may be found not liable when a downstream user infringes. The wording is “may”, and the drafters reportedly softened it from a more absolute draft, so courts keep discretion.

Patents. Under Article 14, an AI-related invention is patentable if it uses technical means that follow natural laws, solves a technical problem and produces a technical effect consistent with natural laws. The exceptions are inventions that violate law or public morality, harm the public interest, or lack a natural person’s substantive contribution. A natural person who makes a creative contribution to the invention’s substantive features is the inventor. In practice, ownership disputes will turn on whether a human can show that contribution, so keep records of prompts, outputs and design iterations.

Data. Article 16 protects lawfully acquired data interests and gives three routes by data type: copyright for compilations that qualify as works, trade-secret law for confidential data, and Article 13 of the Anti-Unfair Competition Law for other data. It also makes fabricating interference data, malicious labeling and adversarial attacks that harm AI system safety actionable.

6. Other provisions to watch

Article 4 treats unauthorized AI face-swapping, voice cloning and digital re-creation of the deceased as personality-rights infringements. Article 8 allows injunctions against AI-enabled personality-rights violations, including orders to service providers. Article 19 requires litigants to verify AI-generated filings and disclose their use, which matters for any foreign party and its counsel litigating in China.

7. What the Opinion leaves open

The Opinion says nothing on whether AI-generated content is copyrightable or on whether training on copyrighted works without permission infringes. At the press conference the SPC said views were sharply divided during drafting and understanding needs to deepen.

The gap will not stay empty. Article 22 lets higher courts take over important and novel AI cases, and Article 22 also points to the case library for building guidance. Early decisions, such as the 2023 Beijing Internet Court text-to-image case, are likely to carry weight beyond their facts. Parties in early AI cases should litigate with that in mind.

8. Four steps to take now

① Build a training-data file. Source, license, cleaning logs and model-version history, ready to produce under Articles 12 and 17.

② Test your complaints channel. It should accept prompt-based evidence of infringement, and you should be able to block related generation quickly.

③ Document open-source releases. State license scope, known functional and safety risks and security testing.

④ Check your personal-data basis. Review whether publicly available personal information used in training stays within the context in which it was published.