Authored by Zhigang Zhu & Paul Ranjard

Precis: The Supreme Court’s first issued document guiding the adjudication of AI-related disputes is set to have a significant practical influence on how both rights holders and the courts approach complex issues, such as the allocation of responsibility and treatment of technical evidence. 

On 7 September 2026, the Supreme People’s Court (SPC) of China published its “Opinion on the Adjudication of Disputes Involving Artificial Intelligence”. It is the first document issued by China’s highest court specifically to guide the adjudication of AI-related disputes.

The Opinion is neither legislation nor a formal judicial interpretation, and it does not create a separate body of AI law. It directs courts to apply existing rules under the Civil Code, Copyright Law, Patent Law, Personal Information Protection Law and other legislation relevant to AI-related disputes. Its practical influence is nevertheless likely to be substantial because it provides more explicit recommendations on some complex issues like the allocation of responsibility and the treatment of technical evidence, particularly where lower courts’ approaches diverge.

The Opinion does not give specific directions over the legality of large-scale training or the copyright status of AI-generated content, which will remain to be adjudicated on a case-by case basis.

Copyright in AI-generated content

The Opinion does not decide when AI-generated content qualifies for copyright protection. The SPC’s explanatory article confirms that this issue was deliberately left open because no sufficient consensus had been reached.

The courts examine the user’s contribution to the specific expression claimed as a work. The number of prompts and their causal relations with the generated content are an important criterion.

In Spring Breeze (Beijing Internet Court, 17 November 2023), the plaintiff Li Yunkai used a platform called Stable Diffusion to create an image of a young woman and published the portrait. Another person, Liu Yuanchun, posted an article using this portrait. Li sued for copyright infringement. The court analysed the full sequence of prompts that Li had recorded and found that the portrait reflected the plaintiff’s personal choices. The image was therefore protected as a work of fine art.

Butterfly Chair (Zhangjiagang Court, Jiangsu province, 19 March 2025) reached the opposite conclusion. Feng, the plaintiff, used a prompt "children's chair with jelly shape of cute pink butterfly - glass texture - light background", to generate an image showing a chair with butterfly wings. He posted the image and the prompt on Red Note. He was, then, approached by Zhu, a chair manufacturer, for the authorization to manufacture the chairs, but no deal was signed. Zhu, then, used the same prompt and the platform produced a similar image. Feng sued for copyright infringement. Her claim was dismissed because it was shown that using the same prompt led to, each time, a different result.

Liability for AI-related infringement

Articles 7 and 12 describe the role of the AI developer, AI service provider, AI user and rights holder.

Article 7 addresses liability of generative-AI service providers. A rights holder may send a notice to the service provider identifying AI-related infringement, supported by preliminary evidence and the rights holder’s identity. The service provider must take measures such as stopping the generation of the infringing content or blocking the prompts that lead to the generation of such content. A provider that fails to act may be held jointly liable with the user for the resulting increase in harm.

Article 12 concerns AI-related IP infringement. Like in any IP case, the rights holder needs to provide evidence of the alleged infringement. If the AI developer claims that no infringement has been committed, it needs to provide evidence concerning the source of its training data, records of the training process and the model’s operating method. A developer cannot expect the court to accept a general statement that the model is a “black box” or that the output was automatically generated. Concerning the AI user who created a work substantially similar to an existing work, its liability may depend on whether they knew, or should have known, of the existence of such prior work.

The above general principles have been implemented by the courts, with sometimes diverging results.

In a case about AI-generated infringement of Japanese superhero Ultraman, adjudicated in March 2024 by the Guangzhou Internet Court, the user of an Internet platform entered prompts relating to Ultraman and the service provider directly produced images that wholly or partly reproduced protected features of the character.

The court considered that the platform had full control over the operation of the AI service. The court also examined the platform’s preventive measures and found that although it had introduced keyword filtering, alternative prompts could still generate substantially similar Ultraman images. The platform had no adequate copyright complaint mechanism, gave insufficient warnings about infringement risks and did not prominently identify the images as AI-generated. The court ruled in favour of the plaintiff.

The facts were different in another Ultraman case adjudicated by the Hangzhou Intermediate Court in December 2024. In that case, the court found that the users were able to add their own low-rank adaptation (LORA) by fine tuning an AI model, allowing them to control each step of the creation. In the absence of evidence of a common intention to infringe, the platform was not treated as a direct infringer. However, it was nevertheless held liable for assisting infringement as the court found it should have known of the risk.

In the Medusa case, which was adjudicated on 29 April 2026 by the Shanghai IP Court, the same type of LORA function was used, but it produced a different result for the platform. A user extracted more than 20 images of the Medusa character from the animated series ‘Battle Through the Heavens’, assembled them as a training pack and used the platform to train two Medusa LORA models. The user then published the models. When used with different prompts, the model consistently generated images containing protected features of Medusa’s face, costume and colouring.

The Shanghai courts held the user liable. The platform was exonerated as it neither duplicated/disseminated the infringing contents nor aided or abetted the user in its infringing act. Unlike Ultraman, the Medusa character was not considered so readily identifiable that a general platform should have foreseen or prevented the user’s infringing act.

Training data

The Opinion also sidesteps the question of whether the unauthorised use of copyrighted works for model training is permissible. The SPC expressly identifies this as another copyright issue on which no consensus has been formed.

Article 24 of China’s Copyright Law specifies circumstances in which a work may be used without permission of or payment to copyright owner, including personal study, appropriate quotation, news reporting and limited classroom teaching or scientific research, provided such use neither disrupts the normal exploitation of the work nor unreasonably harms the rights holder’s legitimate interests. AI training is not expressly included.

The first-instance judgment in the Hangzhou Ultraman case contains a discussion under which training may qualify as fair use where the purpose is to analyse patterns and features rather than reproduce protected expression, and where the training does not affect normal exploitation or unreasonably harm the rights holder.

Evidence and disclosure

The Opinion gives technical evidence a larger role in AI litigation.

Article 17 provides that where a party controls documents or electronic data and refuses to produce them without justification, the court may accept the opposing party’s account of what that evidence would attest.

Article 18 sets out factors for examining AI-generated evidence. These include:

• the design of the prompts;
• their effect on the output;
• similarity to the protected work;
• consistency across repeated tests;
• the training method;
• the algorithm design; and
• the operation of content filters.

Article 19 separately requires parties to verify AI-generated pleadings and case research reports before submitting them to a court, disclose the use of AI assistance and assume responsibility for the accuracy of the material.

Patents, open source and data

The remaining IP provisions are narrower.

Article 14 confirms that an AI-related invention may be patentable if it uses technical means, solves a technical problem and produces a technical effect consistent with natural laws. A natural person must still have made a creative contribution to the substantive features of the invention to be named as inventor.

Article 16 does not create a general property right in data. It protects particular data or datasets through existing legal routes. A dataset may be protected by copyright law if it qualifies as a compilation or other type of work, or under the Anti-Unfair Competition Law where it constitutes trade secret. The provision also covers manipulation of data, malicious labelling and adversarial attacks that damage the safe operation of AI systems.

Takeaways

Given the numerous references made by the Opinion to existing laws and regulations, it appears that the SPC opts to toe the line in such a complex and technical field, and prefers to guide the people's courts in the best way to adjudicate AI-related IP cases: by applying existing laws.