Background of the Hangzhou Case

The defendant runs a platform that offers Checkpoint base models and LoRA (Low-Rank Adaptation) models, enabling functions like image generation and online model training. When users search for Ultraman-related keywords on the platform's homepage, they find AI-generated images and Ultraman LoRA models. Users can save, apply, publish, or share these images and models through links. The Ultraman LoRA model is created when users upload Ultraman images, select a base model, and adjust training parameters. Other users can then input prompts, choose the base model, and apply the Ultraman LoRA model to generate images that closely resemble the Ultraman character.

Plaintiff's Perspective

The plaintiff argues that the defendant has made infringing images and models available on the information network, violating copyright laws regarding information dissemination. The AI platform's generative technology, particularly in training the Ultraman LoRA model, is said to repeatedly produce infringing images, undermining the competitive rights the plaintiff has built over time and constituting unfair competition. Consequently, the plaintiff seeks a court order for the defendant to stop the infringement and to compensate for economic losses and reasonable expenses amounting to 300,000 yuan.

Defendant's Defense

The defendant claims that the AI platform utilizes third-party open-source model codes tailored to its operational needs and does not provide training data, thus qualifying for the safe harbor provision that exempts it from liability, and therefore does not constitute infringement.

Court's Perspective

The Hangzhou court's analysis is more detailed, differentiating various application scenarios and defining infringement liability. Unlike the Guangzhou court, which directly labeled the platform as infringing, the Hangzhou court concluded that while the defendant offers generative AI services, the infringing training materials are uploaded by users, who also decide to publish or share the infringing models and images. There is no evidence that the defendant collaborated with users to create infringing works, nor did the defendant directly engage in actions governed by the right of information dissemination, thus not constituting direct infringement.

The court also found no evidence of unfair competition: the LoRA model training service and reference image service are designed to broaden the application scenarios and functionalities of generative AI, offering users more personalized content customization to enhance creative efficiency without breaching good faith and business ethics. From a market competition perspective, technology itself is neutral. If users create generative AI content in line with the platform's service agreement while respecting intellectual property rights, it does not infringe upon copyright holders' legitimate rights or harm consumer rights and public interests.

The Hangzhou court primarily determined that the defendant engaged in indirect infringement, stating that the defendant company should have known that network users were using its network services to infringe upon the right of information network dissemination and failed to take necessary measures, thus not fulfilling its reasonable duty of care, resulting in aiding infringement. This conclusion is based on the presence of multiple Ultraman anime images in the Works and Overlay Model LoRA sections on the platform's homepage. The Overlay Model LoRA section includes categories like IP Works, allowing users to select this category when publishing LoRA models, and the cover or example images of these models prominently display infringing content, making it easily noticeable.

Legal Insights

AI presents new legal challenges, and court rulings can vary significantly across regions. Additionally, China does not follow a case law system; without relevant cases being included in the People's Court database or recognized by the Supreme People's Court, they cannot serve as definitive solutions to related legal issues.

In discussions about the two Ultraman AI cases in Guangzhou and Hangzhou, media attention often centers on the training materials, but the core issue is more about keyword filtering and the platform's responsibilities regarding user-generated content (UGC), which are not directly tied to training data compliance. The cases reviewed by the Guangzhou and Hangzhou courts are quite similar, with the Guangzhou court asserting that any output of copyright content constitutes infringement. If this reasoning is accepted, AI platforms would face significant infringement risks; any AI capable of generating copyright content would be deemed infringing, necessitating content filtering in both input and output environments. Conversely, the Hangzhou court maintains that merely outputting copyright content does not equate to infringement; it is UGC, and only specific sections created by the platform constitute infringement.

The Hangzhou court's determination of platform infringement hinges on the presence of multiple Ultraman anime images in the Works and Overlay Model LoRA sections. The Overlay Model LoRA section allows users to select IP Works when publishing LoRA models, and the cover images directly display infringing content, making it easily identifiable. Given the vast number of copyrighted works, it is nearly impossible for the platform to preemptively filter all copyrighted content, as this would require prior knowledge of all copyrighted works, which is an unreasonable expectation.

Essentially, the Guangzhou court conflates the ability to generate infringing works with actual infringement, overlooking that it was the plaintiff, as a user, who prompted the AI platform to create infringing works. Without the plaintiff's active request, the platform would not have generated Ultraman images. Article 7, Paragraph 2 of the Supreme People's Court's Provisions on Evidence in Civil Litigation Involving Intellectual Property Rights states that evidence of infringement based on the actions of others can be used by the rights holder to sue for infringement, except when the alleged infringer's actions are solely based on the rights holder's evidence-gathering efforts. The Ultraman Case reviewed by the Guangzhou court essentially falls under this exception.

Compliance Notes

AI platforms that generate text, images, or videos must be vigilant regarding copyright content generation and establish an intellectual property response mechanism based on the notice-and-takedown rule. They should avoid recommending, organizing, or ranking copyright content or content involving portrait rights. User agreements regarding infringement liability should be clearly defined, and reminders about generation should be provided through pop-ups, bold text, checkboxes, or other methods.

The keyword filtering mechanism should not only focus on content compliance but also consider intellectual property compliance. It is crucial to maintain user logs, ideally retaining them for three years in line with the statute of limitations, to present electronic evidence in defense against fishing for evidence.