Shanghai IP Court Issues Final Ruling in City’s First AI Model Copyright Case
Published 11 May 2026
Yu Du
On 29 April 2026, the Shanghai Intellectual Property Court issued its final ruling in Shanghai’s first copyright dispute involving a large-scale AI model. The case addressed a novel legal issue at the intersection of generative AI and copyright law: whether training and publishing AI models capable of reproducing an anime character constitutes copyright infringement.
The dispute centered on the character “Medusa” from the popular Chinese animated series Battle Through the Heavens.
Background
The plaintiff, Company Y, holds the copyright to the Medusa character. The dispute arose after an individual user, Mr. Li, collected screenshots of the character and used them to assemble a training dataset. Using an AI image-generation platform operated by Company H, he subsequently trained and publicly released two Medusa-themed “LoRA models”.
These models enabled other users on the platform to generate images closely resembling the original character design.
Company Y sued both Mr. Li and Company H. Against Mr. Li, it alleged infringement of its rights of reproduction, adaptation, and communication through information networks, as well as acts of unfair competition. Against Company H, the plaintiff argued that the platform operator had failed to fulfill its duty of care by allowing the models to be distributed.
At first instance, on 3 November 2025, the Shanghai Jinshan District People’s Court ordered Mr. Li to cease the infringing conduct and pay RMB 50,000 (USD 6,900) in damages and legal expenses. However, the court rejected the remaining claims, including platform liability and unfair competition. (For details of the first-instance proceedings, please see our previous article: Shanghai Court Issues First Instance Judgment on AI Model Infringement) Both parties appealed, leading to the final appellate ruling issued on 29 April 2026.
Key Findings of the Court
The Shanghai Intellectual Property Court upheld the first-instance judgment in full and clarified several important legal principles concerning AI-generated content and platform liability.
1. AI Training Does Not Automatically Constitute Reproduction
The court distinguished AI model training from traditional acts of copying.
According to the court, merely inputting copyrighted images into a model training process does not automatically infringe the copyright holder’s reproduction right. The court indicated that infringement analysis should focus primarily on whether the model-generated output substantially reproduces protected expression of the original work and makes that expression accessible to the public.
In other words, the legal focus should be placed on the infringing output rather than the technical training process itself.
2. Platform Operators Are Not Automatically Liable
The court further held that Company H, as a technology service provider, did not directly participate in copying or disseminating the infringing works.
Under Chinese copyright law, joint or contributory infringement requires evidence of intent or fault. The court found no such evidence in this case. It also examined whether the platform had fulfilled its duty of care and noted that Company H had implemented several compliance measures, including:
user reporting mechanisms; pre-publication review procedures; and prompt takedown actions upon receiving a valid complaint.
Given the technical complexity of LoRA models and the fact that the alleged infringement would not necessarily be obvious to an ordinary reviewer, the court concluded that Company H had acted reasonably and therefore should not bear joint liability.
The ruling signals that Chinese courts may take a relatively protective approach toward AI platforms that maintain effective compliance and notice-and-takedown mechanisms.
3. “Medusa” Was Not Protected as a Commercial Identifier
The court also rejected Company Y’s unfair competition claims.
It held that “Medusa” primarily refers to a mythological figure with broad public recognition, rather than functioning uniquely as a commercial identifier associated with Company Y. Because the plaintiff had not demonstrated that the name itself had acquired distinctiveness as a source identifier in commerce, the claim did not qualify for protection under China’s unfair competition laws.
The decision illustrates the difficulty of claiming exclusive rights over widely used or culturally established character names without strong evidence of secondary meaning.
4. Damages Award Upheld
Finally, the court upheld the RMB 50,000 damages award against Mr. Li.
In determining the amount, the court considered factors including:
the popularity and commercial value of the Medusa character; the defendant’s degree of fault; the scope and duration of the infringing conduct; and the public dissemination of the LoRA models.
Comments from the Presiding Judge
Judge Liu emphasized that courts must strike a balance between robust copyright protection and the need to preserve room for technological innovation.
He noted that AI training differs fundamentally from conventional copying because large-scale models generally learn statistical relationships and patterns rather than storing or reproducing original works in a direct, one-to-one manner.
Accordingly, infringement should be assessed based on whether the generated output reproduces the copyrighted work’s protected expression, rather than treating every act of model training as inherently unlawful.
Judge Liu also stressed that technology providers should not face automatic liability merely because their systems are capable of misuse. Liability should arise only where a provider intentionally participates in infringing conduct or knowingly fails to take reasonable preventive measures.
He concluded by acknowledging that AI-related copyright disputes remain an evolving area of law and that further legislative and judicial guidance will likely be necessary to balance innovation with effective rights protection.
Comment
This ruling provides important guidance for both copyright owners and AI platform operators in China. For AI companies, the decision underscores the importance of maintaining robust compliance mechanisms. Platforms that implement good-faith safeguards — including reporting systems, review procedures, and prompt takedown measures — may substantially reduce their exposure to secondary liability. At the same time, the case confirms that copyright holders may still pursue users who intentionally train and distribute infringing AI models.
More broadly, the decision reflects an emerging judicial approach in China: copyright law remains fully applicable in the AI era, but liability will generally attach to the party responsible for generating or disseminating infringing outputs, rather than to the underlying technology itself, absent evidence that the platform knowingly facilitated or wilfully ignored the infringement.
The dispute centered on the character “Medusa” from the popular Chinese animated series Battle Through the Heavens.
Background
The plaintiff, Company Y, holds the copyright to the Medusa character. The dispute arose after an individual user, Mr. Li, collected screenshots of the character and used them to assemble a training dataset. Using an AI image-generation platform operated by Company H, he subsequently trained and publicly released two Medusa-themed “LoRA models”.
These models enabled other users on the platform to generate images closely resembling the original character design.
Company Y sued both Mr. Li and Company H. Against Mr. Li, it alleged infringement of its rights of reproduction, adaptation, and communication through information networks, as well as acts of unfair competition. Against Company H, the plaintiff argued that the platform operator had failed to fulfill its duty of care by allowing the models to be distributed.
At first instance, on 3 November 2025, the Shanghai Jinshan District People’s Court ordered Mr. Li to cease the infringing conduct and pay RMB 50,000 (USD 6,900) in damages and legal expenses. However, the court rejected the remaining claims, including platform liability and unfair competition. (For details of the first-instance proceedings, please see our previous article: Shanghai Court Issues First Instance Judgment on AI Model Infringement) Both parties appealed, leading to the final appellate ruling issued on 29 April 2026.
Key Findings of the Court
The Shanghai Intellectual Property Court upheld the first-instance judgment in full and clarified several important legal principles concerning AI-generated content and platform liability.
1. AI Training Does Not Automatically Constitute Reproduction
The court distinguished AI model training from traditional acts of copying.
According to the court, merely inputting copyrighted images into a model training process does not automatically infringe the copyright holder’s reproduction right. The court indicated that infringement analysis should focus primarily on whether the model-generated output substantially reproduces protected expression of the original work and makes that expression accessible to the public.
In other words, the legal focus should be placed on the infringing output rather than the technical training process itself.
2. Platform Operators Are Not Automatically Liable
The court further held that Company H, as a technology service provider, did not directly participate in copying or disseminating the infringing works.
Under Chinese copyright law, joint or contributory infringement requires evidence of intent or fault. The court found no such evidence in this case. It also examined whether the platform had fulfilled its duty of care and noted that Company H had implemented several compliance measures, including:
user reporting mechanisms; pre-publication review procedures; and prompt takedown actions upon receiving a valid complaint.
Given the technical complexity of LoRA models and the fact that the alleged infringement would not necessarily be obvious to an ordinary reviewer, the court concluded that Company H had acted reasonably and therefore should not bear joint liability.
The ruling signals that Chinese courts may take a relatively protective approach toward AI platforms that maintain effective compliance and notice-and-takedown mechanisms.
3. “Medusa” Was Not Protected as a Commercial Identifier
The court also rejected Company Y’s unfair competition claims.
It held that “Medusa” primarily refers to a mythological figure with broad public recognition, rather than functioning uniquely as a commercial identifier associated with Company Y. Because the plaintiff had not demonstrated that the name itself had acquired distinctiveness as a source identifier in commerce, the claim did not qualify for protection under China’s unfair competition laws.
The decision illustrates the difficulty of claiming exclusive rights over widely used or culturally established character names without strong evidence of secondary meaning.
4. Damages Award Upheld
Finally, the court upheld the RMB 50,000 damages award against Mr. Li.
In determining the amount, the court considered factors including:
the popularity and commercial value of the Medusa character; the defendant’s degree of fault; the scope and duration of the infringing conduct; and the public dissemination of the LoRA models.
Comments from the Presiding Judge
Judge Liu emphasized that courts must strike a balance between robust copyright protection and the need to preserve room for technological innovation.
He noted that AI training differs fundamentally from conventional copying because large-scale models generally learn statistical relationships and patterns rather than storing or reproducing original works in a direct, one-to-one manner.
Accordingly, infringement should be assessed based on whether the generated output reproduces the copyrighted work’s protected expression, rather than treating every act of model training as inherently unlawful.
Judge Liu also stressed that technology providers should not face automatic liability merely because their systems are capable of misuse. Liability should arise only where a provider intentionally participates in infringing conduct or knowingly fails to take reasonable preventive measures.
He concluded by acknowledging that AI-related copyright disputes remain an evolving area of law and that further legislative and judicial guidance will likely be necessary to balance innovation with effective rights protection.
Comment
This ruling provides important guidance for both copyright owners and AI platform operators in China. For AI companies, the decision underscores the importance of maintaining robust compliance mechanisms. Platforms that implement good-faith safeguards — including reporting systems, review procedures, and prompt takedown measures — may substantially reduce their exposure to secondary liability. At the same time, the case confirms that copyright holders may still pursue users who intentionally train and distribute infringing AI models.
More broadly, the decision reflects an emerging judicial approach in China: copyright law remains fully applicable in the AI era, but liability will generally attach to the party responsible for generating or disseminating infringing outputs, rather than to the underlying technology itself, absent evidence that the platform knowingly facilitated or wilfully ignored the infringement.