Skip to content 
Abstract
The rapid growth of generative artificial intelligence has created acute doctrinal challenges for copyright law. Courts across major jurisdictions are being forced to determine whether acts such as dataset creation, model training, model dissemination, and AI-generated outputs constitute copyright infringement.
This article examines three emerging judicial approaches: (1) the 2025 first-instance judgment of the Shanghai Jinshan District People’s Court in Yuewen Group v [Defendant] (case number not yet published), (2) the High Court of England and Wales’ decision in Getty Images (US), Inc & Ors v Stability AI Ltd [2025] EWHC 2863 (Ch), and (3) recent judgments from the Regional Court of Hamburg (308 O 42/23) and the Regional Court of Frankfurt (2-06 O 175/23). Together, these decisions illustrate sharply divergent views on the legality of AI training and the allocation of copyright liability.
1. Introduction to AI Training Copyright
The emergence of generative AI systems capable of producing images, text, and audio has fundamentally disrupted conventional copyright assumptions grounded in human authorship. Courts are now confronted with questions such as whether the ingestion of copyrighted material for training constitutes reproduction; whether a model’s latent parameters may themselves represent infringing copies; whether disseminating a trained model infringes dissemination rights; and whether outputs that resemble protected works should trigger liability.
2. China: Yuewen Group v [Defendant] (Shanghai Jinshan District People’s Court, 2025)
2.1 Background
The claimant is 阅文集团, a well-known Chinese media/IP-development company. The case is described as the first case in Shanghai involving alleged copyright infringement by an AI large-model (or LoRA-style) generated output — i.e. AI-driven generation that allegedly re-uses prior copyrighted material. The allegedly infringed work appears to be a character image from the IP series 斗破苍穹 — specifically a character (or “image”) named “美杜莎” belonging to that IP.
The defendant reportedly took a collection of existing images of that character (20+ images), used them to train a LoRA-style model, then published that model. Other users of the model could then generate new images using prompts, producing images substantially similar (or allegedly identical in expressive content) to the original IP character.
The court (上海市金山区人民法院) in its first-instance judgment ruled in favour of Yuewen — found that using pre-existing copyrighted character images to train an AI model, then making that model available so that it can generate new images substantially similar (or reproducing the original’s expressive elements), amounts to copyright infringement (at least of certain exclusive rights like reproduction and dissemination).
2.2 Significance
The Chinese judgment provides a comprehensive and rights-holder-oriented approach. The Court treated the entire pipeline — training, model release, and output generation — as within the scope of copyright protection, especially when the generated outputs reproduce key expressive elements of existing works.
The court reportedly did not find (or at least did not base its ruling on) an infringement of “adaptation rights” (i.e. “改编权”) in this case. In other words, the court did not conclude that the AI-generated images constitute a new derivative work under the original work’s “adaptation” right, because the AI-generated images did not themselves reflect independent “substantial human intellectual input,” they did not qualify as “works” in the sense of the Copyright Law — at least for purposes of granting them new copyright protection. This suggests a continuing legal and doctrinal tension: while copying and distribution via AI-generated outputs can be infringing, not all AI-generated works will themselves be considered new “works” capable of their own copyright.
3. United Kingdom: Getty Images v Stability AI [2025] EWHC 2863 (Ch)
3.1 Background
Getty Images is a leading global visual content creator, licensing millions of high-quality images and videos. Stability AI, founded in 2019, specializes in machine learning software, including the Stable Diffusion model. Stable Diffusion is a generative AI model that transforms user prompts into images, trained on large datasets, including LAION-5B.
Getty Images brought an action against Stability AI, alleging that Stability trained its generative-AI model — Stable Diffusion — using millions of Getty’s copyrighted images without authorisation, and that by making the model available (or allowing its use) the company infringed Getty’s copyright, trade-mark and related rights. So Getty Images claims primary and secondary copyright infringement, database right infringement, trademark infringement, and passing off.
Stability AI claims that the training did not occur in the UK, and the model does not store training data, focusing instead on inference. While Stability admits some LAION-Subsets contain Getty Images’ URLs, it disagreed on the number of Getty Images used in training. Stability also argued that filters may be applied to remove undesirable images.
3.2 Significance
The High Court’s decision is notable for its restraint. It declined to make findings on the legality of AI training itself because Getty was unable to show that the training occurred in the United Kingdom; the servers used were located abroad, leaving the key acts outside UK jurisdiction. As a result, the Court did not address the broader question of whether training on copyrighted works constitutes infringement under UK law.
However, the Court did consider the nature of the trained model. It rejected the argument that the Stable Diffusion model weights contained copies of Getty’s photographs. The Court held that a latent diffusion model does not store images in a form that satisfies the UK’s definition of a “copy.”
The one area where infringement was found concerned trade marks. Some AI-generated images reproduced distorted versions of Getty’s watermark, constituting trade mark infringement under the Trade Marks Act 1994.
The UK’s approach stands in marked contrast to China’s. The High Court avoided embracing or rejecting copyright liability for AI training and emphasised territorial limits. By holding that a trained diffusion model does not embody infringing copies, the judgment suggests a relatively cautious — even conservative — approach to regulating AI development. Liability was confined to a narrow instance of trade mark misuse.
Although the case leaves many issues unresolved, it signals that UK courts may be reluctant to characterise AI models themselves as infringing reproductions.
4. The European Union: Hamburg and Frankfurt Divergences
4.1 The Hamburg Judgment (308 O 42/23)
The Hamburg court considered whether compiling a dataset of copyrighted works for possible AI training fell within the EU’s text and data mining (TDM) exceptions under Directive (EU) 2019/790. It held that dataset creation is permissible where the user has lawful access and rights holders have not opted out. The judgment deliberately avoided addressing downstream model training or commercial deployment. Consequently, the safe harbour is narrowly circumscribed.
4.2 The Frankfurt Judgment (2-06 O 175/23)
Frankfurt addressed the legality of training a commercial AI language model on copyrighted song lyrics. The Court held that such training fell outside the TDM exceptions and infringed the reproduction right. Because the model could reproduce protected lines of lyrics, infringement was established. The developer was required to cease using the lyrics as training data.
4.3 Significance
The EU jurisprudence reveals a fragmented doctrinal landscape. Hamburg recognises a narrow TDM carve-out limited to dataset compilation, while Frankfurt adopts a more rights-protective stance in relation to commercial training and output reproduction. The EU’s legal structure, heavily shaped by the Copyright Directive and TDM regime, complicates the development of a cohesive approach.
5. Comparative Analysis Across Jurisdictions
The three jurisdictions diverge sharply in their treatment of AI training and liability. China adopts a holistic approach that treats the entire AI pipeline — training, dissemination, and output — as potential infringement. This approach extends existing copyright rights without relying on specialised statutory frameworks.
The UK, by contrast, takes a circumscribed path, emphasising territorial limits and declining to treat diffusion model weights as infringing copies. Its approach is therefore more developer-friendly and doctrinally cautious, with the only infringement identified relating to trade mark misuse rather than copyright.
The EU finds itself between these positions but without internal coherence. Hamburg carves out a permissive approach to dataset creation, while Frankfurt limits the availability of TDM exceptions and subjects commercial training to traditional infringement analysis. The result is a patchwork rather than an integrated system.
Another point of divergence concerns the status of AI-generated outputs. China explicitly denies such outputs the status of “works.” The UK has not yet addressed the question, and the EU courts have only implied that outputs may infringe where they reproduce protected material, without addressing their originality.
6. Conclusion
The Chinese, UK, and EU decisions examined in this article illustrate the lack of international alignment on how copyright law should respond to the realities of AI training. China has moved swiftly towards a rights-holder-centered model, the UK remains markedly cautious and doctrinally conservative, and the EU’s jurisprudence remains fractured and heavily influenced by statutory TDM exceptions.
As generative AI becomes increasingly embedded across creative and commercial sectors, courts and legislators will need to confront unresolved questions about authorship, reproduction, and the permissibility of training on copyrighted material. Harmonisation appears unlikely in the near term, but these early cases will shape the emerging global debate on AI and copyright.
If you would like some further information, please get in touch with our team today info@lawdit.co.uk. We are able to provide legal services in both Chinese and English.
