Issue involved:
Whether the provision of image-generation models by a generative artificial intelligence (AI) platform constitute infringement of the right of communication through information networks?
Mengchun’s Comments:
- To foster the development of generative AI technology, a more lenient and inclusive standard should be applied when assessing whether data input and training activities constitute copyright infringement. However, a stricter standard should be used when evaluating the output and use of generated content.
- Data training activities should generally be considered “transformative use” and may qualify as “temporary reproduction“, thus falling within the scope of fair use.
- Generative AI service providers are not obligated to conduct prior reviews of user-input data, such as prompt texts and training images, or the dissemination of generated content. Liability for contributory infringement only arises when the provider is aware of or actively contributes to specific infringing acts.
- When determining contributory infringement, courts should consider multiple factors, including: nature of the generative AI service (e.g. whether it is based on open-source models or is tailored for specific applications); monetization model (whether the platform directly profits from the generative services it provides); notoriety of the copyrighted works; obviousness of infringement; foreseeability of infringement risks (whether the defendant should have reasonably foreseen the possibility of infringement); and platform’s preventive measures (whether they took proactive and reasonable measures to prevent infringement).
Summary of the First Instance Judgment:
1. Background
- Shanghai Character License Administrative Co., Ltd (hereinafter referred to as the “Plaintiff”) holds the exclusive copyright license for the “Ultraman” series of characters in China.
- Hangzhou Jellyfish Intelligent Technology Co., Ltd (hereinafter referred to as the “Defendant”) operates a generative AI platform called “Acgnai” (hereinafter referred to as the “generative AI Platform”), which provides functions such as text-to-image generation, image-to-image generation, reference-based image generation, and model training.
- The generative AI Platform allows users to upload “Ultraman” images to train corresponding LoRA models. It also displayed generated “Ultraman” images and the “Ultraman LoRA model” on its homepage, as well as under its “Recommended” and “IP Works” sections. Users could use this model to generate images that are substantially similar to the “Ultraman” characters and publish or share them on the platform.
Pic：A screenshot of uploading “Ultraman” images to train corresponding LoRA models by Users
Pic: Images generated by the generative AI platform
About LoRA: the core idea of LoRA (Low-Rank Adaptation) is to identify and leverage the redundancy within the parameter matrices of large models. By applying low-rank matrix approximation on top of the original model, LoRA removes information irrelevant to a specific task while retaining essential knowledge for rapid adaptation to that task. This enables LoRA to maintain the original model’s capabilities while efficiently extending its applicability in specific domains
2. Classification of Dispute Acts
- The court divides the disputed actions into three types:
- Training the Model Using Ultraman Images.
- Using Infringing Images as Cover and Example Images for the Model.
- Generating and Distributing Infringing Images Using the Model.
The court ruled that Act 1 constitutes fair use (as detailed in Section 6). However, Acts 2 and 3 were found to infringe upon the Plaintiff’s right to disseminate Ultraman artwork via information networks, applying the “Access + Substantial Similarity” standard.
3. Distinction Between Direct and Contributory Infringement
- Direct Infringement: A generative AI platform may be directly liable if it engages in acts controlled by copyright holders. However, in this case, there is no evidence that the Defendant jointly provided infringing works with users or directly engaged in infringing acts.
- Contributory Infringement: Users upload training materials at the input stage and determine whether to publish or share the generated models and images. As a network service provider, the Defendant may only be held liable for contributory infringement if it can be proven that it had knowledge of or contributed to specific infringing acts.
4. Factors in Determining Contributory Infringement
- Nature of Service and Profit Model: Compared to providers of open-source models, the generative AI Platform directly engages in commercial activities and generates revenue through user membership fees and credit purchases. It also encourages users to publish trained models. Given its business model, the generative AI platform is expected to maintain awareness of content in specific application scenarios and fulfill its duty of care.
- Fame of the Copyrighted Works and Obviousness of infringement: The “Ultraman” series is widely recognized. The generative AI Platform prominently displayed Ultraman-generated images as covers and example images for the Ultraman LoRA model, placing them in easily noticeable locations on the platform. This made the infringing content readily perceptible.
- Predictability of Infringement Consequences: While generative AI platforms typically lack the ability to preemptively detect or intervene in user-generated content, the use of the Ultraman LoRA model increased the AI Platform’s ability to identify and manage infringing content. Additionally, the platform’s features facilitated the repeated use and dissemination of infringing images and models, amplifying the infringement.
- Platform’s Measures to Prevent Infringement: The Defendant’s user service agreement stated that it did not review uploaded or published content. However, after being notified of the lawsuit, the Defendant implemented measures such as blocking related content and conducting intellectual property reviews in the backend. This indicates that the Defendant had the capability to take preventive action but failed to do so in a timely manner, despite prevailing technological standards at the time.
- Conclusion: the court determined that the Defendant should have been aware that users were utilizing its network services to infringe upon the Plaintiff’s right of communication through information networks. However, the Defendant failed to take necessary preventive measures, demonstrating subjective fault. As a result, the court held the Defendant liable for contributory infringement.
5. AI Platforms’ Duty of Care and Ineligibility for Safe Harbor Protection
- Safe Harbor Principle: This principle applies to traditional internet service providers (ISPs) that merely offer hosting services without generating content. Under the “notice-and-takedown” mechanism, ISPs must promptly remove infringing content upon receiving a valid notice from a rights holder to avoid liability for damages.
- AI Platforms’ Heightened Duty of Care: Unlike traditional ISPs, generative AI platforms function as both technology service providers and content suppliers. Users actively contribute to and control the content generated, making these platforms more than passive intermediaries. Given their advanced technological capabilities, AI platforms should bear a heightened duty of care that exceeds that of traditional ISPs and should not automatically qualify for safe harbor protection to avoid liability.
6. Applicability of Fair Use in Data Training
- Transformative Nature of Data Training: The act of data training is considered “transformative” and the reproductions made during data training are temporary, further supporting its classification under fair use.
- Impact on Copyright Holders: In this case, there is no evidence to show that the data training was aimed at using the original works’ creative expression, affecting the normal exploitation of the original works, and unreasonably harming the legitimate interests of copyright holders. Therefore, data training constitutes fair use.
- Conclusion: The court concluded that data training temporarily retains prior works solely for structural feature analysis without displaying them to the public. Even if such use involves copyrighted materials, it qualifies for a fair use exemption. As a result, the court rejected the Plaintiff’s broad request for the Defendant to delete all materials and related data associated with the copyrighted works.
Court:
Hangzhou Internet Court (First Instance)
Hangzhou Intermediate People’s Court (Second Instance)
Final Decision Date:
December 30, 2024
Case No.:
First Instance:
(2024)浙0192民初1587号
(2024) Zhe 0192 Min Chu No. 1587
Second Instance:
(2024)浙01民终10332号
(2024) Zhe 01 Min Zhong No. 10332
