China’s Supreme Court Issues AI Dispute Guidance, Leaving Training Copyright Unsettled

AI companies in China may have to account for how their models were trained and operated when copyright disputes reach court. But the country’s top court has stopped short of deciding whether copyrighted works can lawfully be used for model training in the first place.

On September 7, China’s Supreme People’s Court issued 24 provisions on handling AI-related civil disputes. The Opinions are judicial policy and adjudicatory guidance rather than a new AI law or a binding judicial interpretation, so disputes remain governed by China’s existing civil-law framework.

Fault depends on control, not merely the machine

The guidance generally keeps a fault-based approach. Model developers and service providers are not strictly liable for every output, but they cannot avoid responsibility simply because a system generated the content automatically. Courts may weigh the system’s autonomy, foreseeable risks, the potential harm, and each party’s ability to prevent or reduce it.

That framework matters when a user deliberately prompts an AI system to infringe personality rights, such as through face swapping or voice cloning. Users can be responsible for intentional conduct, while a platform’s liability turns on foreseeability, technical control, and its safeguards. After a rights holder provides verified identity information and initial evidence of infringement, a platform that fails to take necessary measures can face liability for additional harm after notice.

Internal records move closer to the courtroom

In a copyright dispute over generated content, a rights holder generally must first show that the content came from the relevant model and is substantially similar to protected work. Courts may then require a developer or provider to explain its training-data sources, training process, model operation, and scientific grounds where necessary.

Records that may become material evidence

• Training-data sources and the process used to train the model.
• Model operation and the scientific basis relevant to the disputed output.
• Documentary or electronic evidence controlled by a party, whose unjustified withholding can support adverse factual inferences.

The central copyright question remains open

The court did not establish uniform rules on two questions: whether AI-generated content qualifies for copyright protection and whether unauthorized use of copyrighted works for training is lawful. The second question reaches beyond a disputed output to the legality of a model’s underlying learning materials, leaving companies and rights holders without a single nationwide judicial rule on that issue.

The Opinions also separate personal-information rules from copyright. Lawfully public personal information may be used for training within a reasonable scope when a person has not expressly objected and the processing does not materially affect that person’s rights and interests. That permission does not automatically authorize use of copyright-protected expression found in the same material.

For now, the guidance creates a clearer route for courts to examine conduct around AI outputs, platform responses, and evidence held by companies. The more consequential question of training-data copyright is left to individual cases, industry practice, and possible future legislation.