With the rapid development of the artificial intelligence industry, AI creation has emerged, widely applied in writing, image, and video generation, leading to numerous intellectual property disputes. On November 3, 2025, the Jinshan District People’s Court in Shanghai made a first-instance ruling in China’s first copyright infringement case involving large AI models, attracting widespread attention in the industry.
Before delving into the specifics of the case, let’s briefly explain the relevant business model. The LoRA model, in simple terms, is a model that adapts to specific tasks by fine-tuning a large AI model with data; some AI image generation platforms specifically provide LoRA training functions, allowing users to upload materials to generate their own LoRA models. LoRA model platforms typically provide computing power and technology for users to train and share LoRA models. After understanding the relevant business model, we will look at the roles played by the user and the large model provider in this case, and how the court determined their respective responsibilities.
1
Basic Case Facts
The plaintiff in this case is the copyright holder of the character “Medusa” from the well-known IP “Dou Po Cang Qiong”; the defendant company operates a leading AI image generation platform in China, which aggregates various types and themes of AI-generated LoRA models, providing users with AI online image generation services based on large models and numerous LoRA models.
The defendant, Li, is a user of the platform who captured over twenty images of the Medusa character from the “Dou Po Cang Qiong” anime series and created a Medusa image pack. Li used the platform’s “train LoRA” function, inputting the Medusa image pack as training material to generate two Medusa LoRA models. After passing the platform’s machine review, Li published the involved Medusa LoRA models on his account. Other ordinary users could generate various images similar to the Medusa character by inputting different prompts when using the involved Medusa LoRA models.
The plaintiff claimed that the name “Medusa” constitutes a product name with a certain influence, and the involved Medusa LoRA model can specifically generate images of the Medusa character. Li’s actions constitute unfair competition and infringe upon his copyright (including reproduction rights, adaptation rights, and information network dissemination rights); he also argued that the platform allowed a large number of infringing models to exist and failed to fulfill its platform responsibilities. The plaintiff requested the court to order both defendants to cease infringement, apologize, eliminate the impact, and demanded the platform compensate for economic losses and reasonable expenses of 2 million yuan, and Li to compensate 200,000 yuan.
Ultimately, the Jinshan District People’s Court made a first-instance ruling: 1. Defendant Li shall cease infringing the reproduction rights and information network dissemination rights enjoyed by the plaintiff company regarding the Medusa character from the “Dou Po Cang Qiong” series; 2. Defendant Li shall compensate the plaintiff company for economic losses of 30,000 yuan and reasonable expenses of 20,000 yuan within ten days from the date the judgment takes effect; 3. The plaintiff company’s other litigation requests were dismissed.
2
Disputed Focus
This case revolves around the copyright ownership and responsibility allocation of AI-generated content, with the following main points of contention:
1. Legal Protection of the Name “Medusa”:The plaintiff claims that the name “Medusa” has influence and should be protected under the Anti-Unfair Competition Law. However, the court believes that the most widely known meaning of the name “Medusa” is the snake-haired monster from Greek mythology, which is not original to the plaintiff. Many comics, novels, games, and plant varieties have characters or names named “Medusa,” so the plaintiff’s claim that “Medusa” is protected as a “product name with certain influence” lacks sufficient basis.
2. Whether Copyright Infringement Occurred:
(1) It is determined that using others’ materials to train AI models constitutes temporary reproduction, infringing reproduction rights, and publicly releasing the “Medusa” model allows users to access the original work’s unique expression at specific times and places, constituting infringement of information network dissemination rights.
The People’s Court believes that defendant Li, for commercial purposes, captured the original work’s “Medusa” image collection and short video materials without permission to train the LoRA model and publish it, allowing users to generate materials substantially similar to the original work, thereby reproducing the original work’s unique expression. This infringes upon the plaintiff’s reproduction rights and information network dissemination rights regarding the “Medusa” work.
(2) AI-generated images without substantial intellectual input do not constitute works and do not infringe upon adaptation rights.
In the absence of evidence showing that a natural person has made substantial intellectual contributions, images directly generated by generative artificial intelligence do not qualify as artistic works under copyright law. In this case, the involved Medusa LoRA model and Medusa AI-generated images do not show substantial intellectual input from defendant Li, thus do not constitute works, and therefore, defendant Li does not infringe upon adaptation rights.
3. Responsibility of the LoRA Model Platform:From the characteristics of the services and technologies provided by the defendant company, it did not directly participate in the material capture, training, publication, and use of the LoRA models. The technology provided for LoRA models and the function of training LoRA models are neutral technologies that promote the development of generative artificial intelligence, so the defendant company is considered a network service provider. Whether the platform constitutes infringement should be comprehensively determined based on subjective and objective factors. In this case, the platform fulfilled its reasonable duty to inform users, established a complaint and reporting mechanism, and a publication review mechanism. Upon receiving the lawsuit, it promptly took down all Medusa LoRA models and updated the filtering keywords in the platform’s review mechanism. After receiving notice from the plaintiff, it promptly notified overseas AI platforms, showing no subjective fault and fulfilling the obligations of “taking necessary measures” and “transmitting notifications,” thus should not be deemed as infringing.
3
Compliance Recommendations
This case clarifies the legal responsibilities of AI platforms and users in the model training process to some extent, providing significant insights. Through the analysis of the case, we summarize the following compliance recommendations:
(1) User Side:Using third-party materials with copyright without authorization may constitute copyright infringement; users should obtain authorization when training LoRA models.
(2) LoRA Model Platform Side:
A. Improve Publication Review and Complaint Reporting Mechanisms:The platform should establish content review mechanisms for high-profile IP names and character names. If suspected infringing names or materials are detected, they can be intercepted or trigger manual review. At the same time, a complaint reporting mechanism should be established to promptly take down infringing models upon receiving user complaints.
B. Clarify User Prompts and Agreement Terms:Before users use the “train LoRA” function, the platform should prominently remind users not to upload infringing materials and require users to confirm their legal rights or authorization over the materials. The user agreement should clearly state that if user-uploaded content infringes upon others’ rights, the platform has the right to delete the content and ban the account, with the infringement liability borne by the user.
C. Continuously Update Compliance Strategies:As AI technology and infringement methods continue to evolve, platform compliance work must keep pace. It is recommended to regularly monitor industry cases and regulatory dynamics, dynamically adjust review strategies, and timely expand the keyword filtering database and improve rules. Through continuous improvement, the platform can ensure risk control while expanding its scale, balancing user experience and copyright protection.
