In recent years, multiple AI platforms such as ChatGPT and DeepSeek have emerged, sparking a wave of AI storms on the internet. The gradual improvement of AI platforms cannot be separated from the "feeding" of a large amount of corpus and images, and the legal issues caused by artificial intelligence models and their generated content have become the "growing pains" that must be faced. Recently, the Hangzhou Intermediate People's Court in Zhejiang Province concluded a case involving an AI platform. The platform has a dual identity as Ultraman. As a well-known anime IP, it has quickly become popular throughout China since its introduction in the 1990s and has become an idol in the hearts of generation after generation of children. The plaintiff, a cultural development company in Shanghai, has exclusive copyright authorization for the Ultraman series works. The defendant, an intelligent technology company in Hangzhou, is the operating entity of an AI platform. Platform users can use the platform to train and share Ultraman AI models by uploading pictures and other means, and can also use Ultraman AI models to generate works that are substantially similar to the plaintiff's Ultraman image. In February 2024, the plaintiff filed a petition against the defendant to the Hangzhou Internet Court, demanding to stop the infringement and compensate for reasonable expenses. In the era of generative artificial intelligence services, the boundary between platform managers and content producers is becoming blurred. This has also become a key issue in the trial of this case. The network services provided by generative artificial intelligence service providers are different from traditional behaviors such as providing network content, providing network storage space services, and providing search link services The first instance judge of this case, Sha Li, told reporters from the Rule of Law Daily that in generative artificial intelligence services, creative behavior needs to be stimulated by user prompts, the overlap between models and users is enhanced, and the service provider's control over the generated content is reduced. Generative artificial intelligence service providers have dual identities as content producers and platform managers, and belong to a new type of network service provider. Returning to the case itself, the plaintiff claims that the defendant should bear direct infringement liability as a content provider. If it does not constitute direct infringement, the defendant should also bear liability for assisting users in uploading content; The defendant claims that it belongs to a network service provider that provides information storage space for users to upload content, and calling third-party models belongs to providing network technology services, which falls within the platform's exemption scope under the "safe harbor" rule, that is, it claims to only bear the "notification deletion" obligation for platform infringing content. How to combine different types of large model platforms to clarify the reasonable duty of care and fault determination rules for application layer generative artificial intelligence service providers, without precedent to follow. Shali candidly stated, "This poses higher requirements and new challenges for our judicial trial." After comparing the infringement behavior, the plaintiff claimed that the accused infringing image was simply modified based on the Ultraman rights work in question, combined with text prompts. The main function is to present the background content of the image according to the instructions of the prompt words. The accused infringing content and the rights work have a high similarity in Ultraman character image, color matching, clothing details, and other aspects, forming a substantial similarity overall. But to confirm the platform's liability for infringement, this is far from enough. It is understood that generative artificial intelligence typically has four important stages, namely data input, data training, content output, and content utilization. The common analysis of AI copyright infringement disputes in the current academic community is to combine the four stages together for discussion, which means that the model cannot generate meaningful results without data training. Therefore, there is a causal relationship between data training and result generation, which can be combined into a comprehensive evaluation of legal actions. From this perspective, the infringement judgment of any link will affect the overall situation. The trial team of the Hangzhou Internet Court advocates different treatment for different stages. Sha Li said: "This can accurately evaluate each stage and provide a clear idea for the identification of the tort liability of the generative AI service provider." Accordingly, the Hangzhou Internet Court found that although the defendant provided generative AI services, the data input end infringement training materials were uploaded by users, and the users decided to publish or share the infringement models and infringement pictures after they were generated in the content output stage. There is no evidence to prove that the defendant and the user jointly provided infringing works, and the defendant did not directly implement acts controlled by the right of information network communication, which does not constitute direct infringement. So, does the defendant constitute aiding infringement, that is, did the defendant take necessary measures and fulfill reasonable duty of care? It is understood that before the case occurred, there were multiple Ultraman LoRA models in the "Overlay Model LoRA" and other columns of the involved platform, which users could choose from. After overlaying and using Ultraman LoRA models, the AI platform involved in the case can stably output images with Ultraman character image features. Shali said: "Due to the convenience of technology, the pictures and LoRA models generated and released by users can be repeatedly used by other users, which has led to the proliferation of infringement. The defendant should bear a higher duty of care for the generation of relevant infringing content, and should anticipate the possibility of infringement." After comprehensive consideration of the nature of the defendant's provision of generative AI services, the development level of AI technology, the platform's profit model, reasonable measures to prevent infringement and other factors, the Hangzhou Internet Court finally found that the defendant should know that network users use their network services to infringe the right of information network communication, but did not take necessary measures, did not fulfill the reasonable duty of care, and had subjective faults, which constituted a helping infringement of the right of information network communication. Classified measures to achieve balance On September 25, 2024, the Hangzhou Internet Court made a civil judgment: First, the defendant immediately stopped infringing the plaintiff's right of information network dissemination of Altman's works involved in the case, including but not limited to immediately deleting the generated and published Altman pictures and Altman LoRA models involved in the case, stopping providing the release and application services of relevant Altman LoRA models, and taking necessary measures to effectively stop infringement; Secondly, the defendant is ordered to compensate the plaintiff for economic losses and reasonable expenses incurred to prevent infringement. In October 2024, the plaintiff appealed to the Hangzhou Intermediate People's Court. The plaintiff believes that in addition to the infringing images identified in the first instance, there are also a large number of models and works on the platform involved that have certain similarities with Ultraman but actually have certain differences. For these models and works that are difficult to be protected by copyright law, the Anti Unfair Competition Law should be applied for regulation. The focus of the second trial is on whether the involved behavior constitutes unfair competition. Hangzhou Intermediate People's Court also conducts phased analysis. From the perspective of data input and training stages, the data sources of generative artificial intelligence not only include the training database input by the platform itself, but also the data input by users during the platform's service process. When service providers provide model services trained by users to the public, there are massive amounts of users from all over the world who "feed" the models with data, and the legality and copyright status of these data may vary. In this situation, if service providers are strictly required to review and verify every piece of data input by users one by one, it is neither feasible nor in line with its legal attributes, which will undoubtedly increase the burden of development supervision and hinder the development of generative artificial intelligence. Therefore, the duty of care of service providers should be commensurate with their identity and information management capabilities Wu Yuanyuan, the person in charge of the second instance of the case and the deputy chief judge of the Intellectual Property Trial Division of Hangzhou Intermediate People's Court, analyzed and said. The judges of Hangzhou Intermediate People's Court have deeper considerations. Wang Jiangqiao, the chief judge of the Intellectual Property Trial Division of Hangzhou Intermediate People's Court, said, "From the perspective of promoting the development of generative artificial intelligence technology, it is advisable to adopt relatively relaxed and inclusive standards for determining infringement of data input and training behavior of large models. For determining infringement of generated content output and usage behavior of large models, relatively strict standards should be adopted, and a balance between development and protection should be achieved through classified measures." Recently, Hangzhou Intermediate People's Court made a second instance judgment, rejecting the appeal and upholding the original judgment. (New Society)