China AI Copyright Law: AI Art, Prompts and Training Data

As of November 4, 2026, whether AI-generated content can be protected by copyright in China is decided case by case under the Copyright Law, whose text does not mention AI, and the Supreme People's Court has left the question open. Beijing and Changshu courts have protected AI images that reflected a person's original intellectual input, while a Zhangjiagang court refused protection for lack of proof of the creative process.

Status as of November 4, 2026: No national rule settles the question. On September 7, 2026, the Supreme People's Court issued an Opinion on trying AI disputes (Fa Fa [2026] No. 10). Court officials said that, because views differ widely, it makes no provision for now on whether AI-generated content can be copyrighted or on unlicensed use of works to train large models. The National Copyright Administration took comments until August 12, 2026 on draft implementing regulations whose text does not mention AI. The Interim Measures for Generative AI Services, in force since August 15, 2023, bar providers from infringing intellectual property rights in training.

Scope: This article covers Chinese law on copyright in AI output, AI training data, platform liability and AI software. It is general information, not legal advice. Only Li v. Liu is described from the judgment itself. The other decisions come from court, government or press reports or a judge's paper, and quotations from Chinese texts are our translations. For a comparative overview, see how AI and copyright differ worldwide.

Can AI-Generated Images Be Copyrighted in China?

Sometimes. The outcome depends on what the person did and what they can prove. The decisions below come from individual local courts, and the Supreme People's Court describes rules on ownership of AI output as something still to be explored.

An article by Supreme People's Court officials explaining the 2026 Opinion says theory and practice are sharply divided on whether AI output can be a work. It records that some Chinese judgments have said yes, and an opposing view that prompts, parameter changes and selection of outputs are not creative acts. The Opinion makes no provision on the point for now.

Li v. Liu: The Beijing Internet Court's AI Image Ruling

The plaintiff generated an image with Stable Diffusion, posted it online, and sued when the defendant used it. The court's translated judgment says he "used prompt words to work on the picture elements such as the character and how to present it, and set parameters to work on the picture layout and composition, which reflects the plaintiff's choice and arrangement."

The court held that the AI model is not the author, because under the Copyright Law an author can only be a natural person or a legal person or unincorporated organization, and that the model's designer is not the author either. It held the plaintiff to be the author and owner of the copyright. It ordered a public apology and 500 yuan for economic loss, and dismissed his other claims.

The court also said a plaintiff in this position should prominently mark the AI technology or model used, in line with good faith and the public's right to know. It accepted his "AI illustration" hashtag as proper practice.

The judgment was a first-instance decision that allowed an appeal to the Beijing Intellectual Property Court. The figure of more than 150 prompts, often repeated elsewhere, does not appear in the court's English judgment text, so this page does not use it.

Why Protection Was Refused: The Butterfly Chair and Prompt Cases

In the Zhangjiagang case reported by Legal Daily, a designer surnamed Feng claimed that a children's chair copied AI-generated images he had posted. He asked for 200,000 yuan. The court said a user must provide original records showing how added prompts and changed parameters adjusted the image, in order to prove individualized choices and a substantial contribution.

Feng had no such records and admitted he could not reproduce the same image, even when he recreated the process in court. The court held the evidence insufficient and the image not a work, and also found no substantial similarity. The judge in the case said content generated only from simple prompts is an idea and not protected expression.

The appeal did not produce a merits ruling. The Suzhou Intermediate People's Court issued a ruling after the appellant failed to pay the appeal fee, and the first-instance judgment took effect.

In the Shanghai case, as reported by CCTV News, a company claimed that its prompts were a written work and that two people infringed by using identical prompts. The Huangpu court held that the prompts were abstract ideas and simple lists of elements, without individualized intellectual input in the expression.

Evidence: What Courts Have Looked At

In the Changshu case, as reported by Suzhou Daily and Legal Daily, the judge reviewed the Midjourney user agreement, logged into the platform in court, and examined the prompt changes and image iterations. The court then limited the copyright to the image itself, so building a physical heart-shaped installation did not infringe. The trial judge said, as Legal Daily reported, that the half-heart shape, the main part edited, was not itself original, and that the court set damages on the facts of the case as a whole.

The Beijing Internet Court published a 2025 case in which a plaintiff claimed a photo that showed signs of being AI-generated. After the third party he named denied his account, he admitted making false statements. The court dismissed his claims and fined him 10,000 yuan, and it says copyright claimants must honestly state whether a work was generated by AI.

The PRC Copyright Law: Articles 3, 11 and 24

The Copyright Law was amended for the third time on November 11, 2020, and the amendment took effect in June 2021. The text we read does not mention AI or algorithms.

Article 3 defines a work as an original intellectual achievement in the literary, artistic or scientific field that can be expressed in a certain form. Its list ends with "other intellectual achievements that meet the characteristics of a work." Article 11 says the natural person who creates a work is its author, and a legal person or unincorporated organization is treated as the author of a work created under its direction, on its behalf and under its responsibility.

Damages under Article 54 follow actual loss, unlawful gains or royalties. Where those cannot be calculated, a court may award 500 to 5,000,000 yuan, and intentional, serious infringement can draw one to five times the amount.

AI Training Data: The Interim Measures and Article 24

Article 7 of the Interim Measures requires providers to carry out training data processing lawfully. They must "use data and foundation models with lawful sources," and where intellectual property is involved they must not infringe rights others hold under law. Article 4(3) separately requires respect for intellectual property.

The Measures apply to generative AI services offered to the public in China. Under Article 2, organizations that develop or apply the technology without offering services to the public in China are outside them. Article 20 lets the national cyberspace authority act against non-compliant services offered into China from abroad.

Article 24 of the Copyright Law lists 13 situations where a work may be used without permission or payment, provided the author and title are named, normal use is not affected, and the author's lawful interests are not unreasonably harmed. We read all 13. None names text and data mining or AI training. They include personal study or research, and limited copying for classroom teaching or scientific research by teaching or research staff without publication.

Item 13 covers "other circumstances provided by laws and administrative regulations," so a list reading does not close the question. The Supreme People's Court's September 2026 Opinion does not decide whether any item covers AI training: Court officials said it makes no provision for now on unlicensed use of works to train large models. Their explanatory article records one view favoring fair use or no infringement, and an opposing view that the fair use provisions leave no room for training.

The Court's 2026-2030 plan tells courts to handle cases on large-model training material prudently. We did not survey judgments on training.

Who Is Liable When AI Output Infringes? The Ultraman Case

A company holding exclusive Ultraman rights sued the operator of an AI platform where users could upload images, train Ultraman LoRA models, and generate substantially similar images. Legal Daily and Daily Business News, the latter reporting a typical case released at a Hangzhou press conference, give the account below. We did not read the judgments.

On September 25, 2024, the Hangzhou Internet Court held that the platform was not a direct infringer, because users uploaded the training material and decided what to publish. It held the platform liable for contributory infringement of the right of network dissemination. Ultraman was very well known, infringing content was visible on the home page, models could be reused, and the platform showed it could act by blocking and reviewing after being sued.

The Hangzhou Intermediate People's Court dismissed the appeal, and the unfair competition claim failed. The head of its IP division said infringement findings on data input and training call for a relatively lenient standard, and output and use a relatively strict one. That is a judge's comment to a newspaper. The case concerned user-trained models on a platform, not a developer's own training.

Item 12 of the Supreme People's Court's September 2026 Opinion now addresses AI output that infringes copyright. Courts are to weigh the type of service, industry features, training data sources, each party's participation, the necessary measures taken and profits when deciding the liability of developers, providers and users.

A developer raising a non-infringement defense shall be ordered to supply training data sources, training process records, the model's operating mode and its scientific basis. A rights holder claiming that a provider infringed through algorithmic technology must supply relevant evidence.

A court shall support a claim against a user who knew or should have known of a prior work and generated a substantially similar work with no reasonable defense.

AI Labeling Rules

The AI-Generated Synthetic Content Labeling Measures were issued by four agencies in March 2025 and took effect on September 1, 2025. They provide for explicit labels and for implicit labels in file metadata, including provider information and a content number. Labeling does not decide who owns the content, although the Li v. Liu court separately said a plaintiff should mark the AI used.

Software Protection for AI Systems

Computer software is a listed type of work under Article 3 of the Copyright Law. Under the Regulations on the Protection of Computer Software, a program's source code and object code are the same work. Article 6 says protection does not extend to the ideas, processes, operating methods or mathematical concepts used to develop software.

Article 15 of the Copyright Law protects a compilation of works, fragments, or data that are not works, where the selection or arrangement of the contents is original. That protection does not allow the compiler to infringe copyright in the original works.

How China Compares with the US and EU

The US Copyright Office concluded in January 2025 that "prompts alone do not provide sufficient human control" given current generally available technology. Li v. Liu credited prompts and parameters, so the two can diverge on similar facts. The Zhangjiagang and Shanghai cases show that Chinese courts also reject thin prompting.

On training, the EU has an express text and data mining exception with an opt-out, as our European Union AI copyright laws page explains. China's Article 24 list names no such exception.

What Changes in 2027

As of November 4, 2026, we found no Chinese law or regulation on AI copyright scheduled to take effect in 2027. That rests on the Supreme People's Court, National Copyright Administration, cyberspace regulator and government pages cited here and on web searches through that date, so it is not a complete survey. We did not read the National People's Congress legislative plan, and unpublished drafts would not appear. Three pending items could change this page:

• Revised Copyright Law implementing regulations: The draft has 49 articles and leaves its effective date blank. The draft text we read, on a government reprint of the notice, does not use the terms for AI, algorithm, training or data mining, though an August 2026 industry forum reported by the Administration raised AI among emerging sectors.
• National Copyright Administration 2026-2030 copyright plan: China Daily reported on September 10, 2026 that the plan calls for developing a fair-use system for AI training data. We read that report, not the plan, and the report gives no rule text or date.
• Supreme People's Court 2026-2030 plan: It directs courts to weigh the specific instructions a person entered and the process of selecting and modifying, and to explore rules on ownership of AI output and the liability of developers, operators and users. The Court's September 2026 release says questions left blank in the Opinion will be addressed when conditions are mature, with no date.

Change Log

• November 4, 2026: Rewritten against official sources. Added the Changshu, Zhangjiagang, Shanghai and Ultraman decisions, the Labeling Measures, the Supreme People's Court plan, its September 2026 Opinion and the draft implementing regulations. Removed the unsupported claims of more than 150 prompts and of a generally output-friendly stance, corrected a link in the opening, and replaced the flat statement that no training exception exists with the actual Article 24 list.
• June 25, 2026: Last full update before this rewrite.

Related articles

• How AI and copyright differ worldwide
• AI copyright law in the United States
• Japan AI copyright laws
• European Union AI copyright laws

Frequently Asked Questions

Is AI-generated art copyrighted in China?

It can be. Courts have protected it where it reflected a person's original intellectual input. The Beijing Internet Court and a Changshu court protected AI images on that basis. A Zhangjiagang court refused protection where the user could not prove the creative process. No national rule settles the question as of November 4, 2026, and the Supreme People's Court's September 2026 Opinion left it open.

Who owns the copyright in an AI-generated image in China?

In Li v. Liu, the Beijing Internet Court held that the user was the author and that neither the AI model nor its designer was. The Copyright Law says the natural person who creates a work is its author, with a separate rule treating legal persons as authors of certain works.

Are AI prompts protected by copyright in China?

A Shanghai Huangpu court held in a first-instance case, reported in March 2026, that the prompts before it were ideas and simple lists of elements, not a work. That decision concerned those prompts, not every prompt.

What records should I keep to claim copyright in AI output in China?

Courts have looked at prompt histories, parameter changes, image iterations and later edits. A Zhangjiagang court rejected a claim where the user lacked original records of the process. There is no official list, and keeping records does not guarantee protection.

Does China have a text and data mining exception for AI training?

Article 24 of the Copyright Law lists 13 exceptions and none names text and data mining or AI training, though the last item covers other situations set by laws and administrative regulations. The Supreme People's Court's September 2026 Opinion left the question open.

What do China's generative AI rules say about training data?

Article 7 of the Interim Measures, in force since August 15, 2023, requires providers offering services to the public in China to use data and foundation models with lawful sources and not to infringe intellectual property rights.

Can an AI platform be liable for infringing images in China?

Yes, on the facts of the Ultraman case. Hangzhou courts held a platform liable for contributory infringement over user-trained models and images because it failed to take necessary measures. The court did not find direct infringement there. Item 12 of the Supreme People's Court's September 2026 Opinion lists factors for allocating liability among developers, providers and users.

Does AI content have to be labeled in China?

The Labeling Measures, in force since September 1, 2025, provide for explicit labels and implicit metadata labels on AI-generated synthetic content. Li v. Liu separately said a plaintiff should mark the AI used, as a matter of good faith.

Is AI-generated content in the public domain in China?

Not automatically, according to the Beijing Internet Court in Feilin v. Baidu, which said software-generated content is not a work but cannot be freely used by the public.