We compare two recent decisions that offer contrasting approaches to how courts treat the issue of copyright and generative AI models: Getty Images v. Stability AI (High Court of Justice, November 4, 2025) and GEMA v. OpenAI (Landgericht München I, November 11, 2025). The analysis is structured around three axes: territoriality, the notion of “infringing copy” and memorization, and the scope of the text and data mining (TDM) exception.
1) United Kingdom – Getty Images v. Stability AI: focus on territoriality and fixation
In the British case, territoriality was decisive: the Copyright, Designs and Patents Act requires that the primary act of infringing reproduction occur within the United Kingdom. As it was not proven that the model had been trained on British territory, the claim was limited to secondary infringement (i.e., distribution/making the model available in the United Kingdom), which also failed because that theory presupposes the existence of a prior infringing copy.
In essence, Judge Joanna Smith held that the weights of the model do not constitute an “infringing copy”: they are the result of learned patterns and do not store protected works; learning is not the same as copying. The court emphasized the absence of literal memorization and that there is no fixation of works in the model or its parameters; hence, the basic requirement of the right of reproduction was not met.
At the same time, the court found limited trademark infringements: in older versions (series 1.x), outputs were generated with watermarks that imitated the Getty and iStock signs, considered historical infringements under s.10(1) and s.10(2) of the Trade Marks Act 1994. However,the isolated nature of these cases and their reduction in later versions were highlighted.
Practical reading (UK): The judgment outlines a strict standard of “copying” that requires the perceptible fixation of content: if the model does not store the protected works, there is no copying for the purposes of copyright. This provides developers with a relatively safe framework if they avoid any literal storage of protected content. Note that in the UK, there are no “TDM” exceptions under European Directive 2019/790, but only a specific exception for non-commercial scientific research (in fact, the basis of Art. 3 of the Directive).
2) Germany – GEMA v. OpenAI: broad notion of reproduction and limits of TDM
GEMA’s lawsuit focused on OpenAI’s unauthorized use of song lyrics during the training of ChatGPT. Here, unlike in the British case, territoriality was not a point of contention (the acts relevant to the court occurred in Germany).
The court did find memorization: the model returned substantial, sometimes almost literal, fragments of lyrics from the GEMA repertoire in response to simple prompts (“what does song X say?”, “how does song X start?”, “give me the lyrics to song X?”). This ability was interpreted as evidence of fixation of the works in the parameters and, therefore, a reproduction in accordance with the broad notion of Article 2 of the InfoSoc Directive and §16 of the German UrhG.
Under EU law, the concept of “reproduction” covers direct or indirect, temporary or permanent copies, in any form, with subsequent perceptibility being sufficient. Therefore, the court indicates that even statistical storage in weights can be reproduction if it allows protected content to be regenerated.
Regarding TDM exceptions, the court distinguished between two phases:
- Phase 1 (preparatory copies for analysis/training): potentially covered by §44b UrhG and Article 4 DSM (text and data mining for commercial purposes).
- Phase 2 (deployment/use of the model): memorization and reuse are not covered because they are no longer copies “necessary” for analysis, but part of the operation/use of the system. Extending the exception by analogy was rejected, and it was affirmed that there is no regulatory gap.
This issue was not widely debated, as OpenAI did not argue in defense of the German transposition of Articles 3 (non-commercial research, not applicable to OpenAI) or 4 of the Directive (commercial uses, albeit subject to limits) – these were not relevant to GEMA’s allegations of infringement in the aforementioned Phase 2. In practical terms, the court emphasized that the TDM exception does not cover the use of the model in Phase 2 when it returns memorized content: TDM exceptions cover the analysis (training) phase, not subsequent reproduction.
3) Comparison of standards and effects
We therefore see two very different approaches: in England, a strict technical approach to the concept of copying has been taken: a perceptible fixation of the original work (used in training) is required. Weights have been considered to be values that abstract patterns and are not equivalent to stored works. In other words, without literal memorization, there is no infringement by reproduction.
In contrast, in Germany, the court has taken a broad interpretation of the concept of reproduction: if the model is capable of regenerating protected content with simple prompts, there is fixation in the parameters and therefore reproduction for legal purposes. The “TDM” exceptions cover preparatory copies, but not memorization or output based on it in Phase 2 of model use.
This contrast anticipates the effects of forum shopping and requires the design of territorially sensitive compliance strategies.
4) Compliance implications and recommendations
If the German approach prevails, companies generating and/or offering general-purpose AI models (GPAI) will need to adopt active compliance combining legal and technical measures:
- Ensure traceability and express licenses for training data, excluding any data without a valid license.
- Conduct periodic tests of memorization risk (strict and indirect).
- In agreements with data/model providers, establish contractual clauses in data licenses that ensure sufficient rights to reproduce and distribute the material (including subsequent availability).
Conclusion
With these two decisions, two standards are proposed: an English view with a stricter concept of copying (and weights are non-infringing abstractions), versus a German approach with a much broader notion of reproduction. Currently, the legal certainty of developers will depend on the territory and the technical-contractual design that minimizes (or discards) memorization and guarantees sufficient licenses on the training material.
Reviewing the many posts and articles on this topic, we note that some of the literature questions whether “memorization” (according to the German court’s view in GEMA) operates as automatic proof of infringement and recalls that copyright protects actual reproduction, not the mere possibility of memorization. At the same time, it is proposed to shift the regulatory focus to outputs (“plagiaristic outputs”), where the risk of infringement could materialize.
This issue is partly reminiscent of the debates on video reproduction machines (the decisions in USA Sony Corp. of America v. Universal City Studios on Betamax, and the defense of substantial noninfringing uses.) and more recently, the decisions on P2P technologies (e.g., MGM Studios, Inc. v. Grokster, Ltd, and the need for a business model and promotion that demonstrate an intention to attract infringing uses).
Therefore, the issue is far from decided!
↳ Discover more content in the section Insight.
↳ Do you want to stay up today about the sector related news? Follow us on LinkedIn.
