Two important AI /copyright decisions handed down in 2025, the Landgericht München I judgment in Gema v OpenAI (reported as LG München I, Endurteil v. 11.11.2025 – 42 O 14139/24) and the U.K High Court’s decision in Getty Images v Stability AI ([2025] EWHC 2863 (Ch)), reach sharply different conclusions on core issues that regularly arise in litigation involving generative AI models, namely:
- whether the training of an AI model that used copyrighted material outside the jurisdiction can give rise to primary infringement claims in the jurisdiction,
- what kinds of evidence are admissible and persuasive to demonstrate how such models function and in particular whether outputs demonstrate memorisation of training inputs, and
- the legal and doctrinal reasoning courts will deploy when deciding these questions.
These differences are not merely doctrinal nuances, instead they bear directly on the legal risk facing developers and deployers of generative models who distribute services globally.
A Short Recap of the Cases and Procedural Matters
The German case Gema v OpenAI
The Munich court (LG München I) heard a claim by a collecting society (GEMA) asserting that certain song lyrics were (i) memorised within OpenAI’s language models and (ii) reproduced in outputs shown to users of a Chatbot. The claim sought injunctive relief, account of profits and damages under the German Copyright Act (UrhG). The defendant OpenAI is US based, but the Chatbot was offered in the EEA through an Irish entity and the service was made available on servers at various EEA locations, including a server located in the Federal Republic of Germany and prompts made in the German language did receive German outputs. The Munich court found for the claimant on core points, holding, inter alia, that memorisation in the model could amount to a reproduction (“Vervielfältigung”) under §16 UrhG and that the operators were liable for infringing outputs because they exercised control (“Tatherrschaft”) over the system.
The English case Getty Images v Stability AI
The English High Court heard a broad claim including primary and secondary copyright infringement, database right, trade mark and passing off allegations. Significantly, prior to and during trial Getty Images abandoned its “Training and Development Claim” because there was no evidence that the relevant training had taken place in the United Kingdom. The claimant also substantially curtailed its “Outputs claim” as the defendant had blocked certain prompts. Mrs Justice Joanna Smith therefore determined the remaining issues on a narrowed factual matrix and accepted that some Getty content had been used in training but concluded (among other things) that (i) the English court could not find primary infringement based on training and development because the evidence did not place those infringing acts in the UK and (ii) for the purposes of the statutory definition of an “infringing copy” or “article” under the CDPA, model weights in current iteration were not themselves an infringing copy. The judgment is careful, technical and heavily informed by expert evidence and statutory construction.
Jurisdiction and Primary Copyright Infringement when Training is Abroad
A focal point of divergence between the two judgments is how each court dealt with the territorial conduit between (a) training that took place (or is conceded to have taken place) outside the jurisdiction, and (b) alleged infringing acts which have effect or manifestation within the jurisdiction.
The German Approach is Functional Presence & Outputs Available to Users in Germany
The Munich judgment treats presence of the service in the EEA and the availability of the model, including server hosting in Germany, as sufficient to support a claim based on reproductions (“Vervielfältigungen“) within the meaning of §16 UrhG where the model had “memorised” the protected works and reproduced them in outputs. The court’s analysis rests on two interlocking propositions:
- Memorisation means reproduction (material fixation): The court concluded that the presence of an original work fixed in the model’s parameters, even if the work is not stored as a contiguous, identifiable file, can constitute a material fixation (“körperliche Festlegung”) and thus a reproduction under §16 UrhG. The court expressly held that it is not necessary for a discrete dataset to be identifiable inside the model and that the decomposition into parameters suffices if those parameters contain the information that enables reproduction.
- Local availability of outputs as indicia of imprint in the model: The Munich court accepted that a straightforward method of proof under §286 ZPO (the court’s powers to draw conclusions on disputed facts) is to compare an original work with an output generated by a simple prompt and that such a comparison can permit the court to conclude that the work was memorised (and therefore reproduced) in the model. The court underlined that, because the Chatbot and the underlying models were made available on servers in the EEA, including in Germany, and produced German outputs, the technical acts giving rise to the memorisation and the accessible outputs fell within its jurisdictional reach for injunctive and other remedies.
Practical effect: The Munich court’s analysis divorces the position of the critical legal act, the “copying” in the sense of fixation, from a formalistic requirement that the training had to be performed physically in the jurisdiction. If the provider makes the model available to German users by hosting servers in Germany, offering German-language outputs and the actual chatbot service in the European Economic Area, the court treated that availability and the demonstrable reproduction of the protected work in outputs as enough to ground remedial jurisdictional effect in Germany.
The English Approach means Territorial Limits and Statutory Construction of the “Infringing Copy” Question
By contrast, the English judgment reached at least two interrelated conclusions that made it difficult for Getty to maintain a primary infringement claim based on training carried out outside the UK:
- Abandonment of the Training & Development Claim: Getty acknowledged that there was no evidence that the training and development occurred in the United Kingdom. As a consequence, it abandoned the “Training and Development Claim”. The court therefore did not have the factual predicate to find that the primary acts of reproducing training data had occurred in the UK.
- Statutory construction of “infringing copy” / “article” (sections 17 and 27 CDPA): The court undertook a detailed statutory analysis and took the view that model weights, in their final iteration, do not themselves store copyright works in the conventional sense, and therefore do not qualify as an “infringing copy” for the purposes of sections 22/23/27 CDPA merely because the training process (which necessarily involved transient reproductions) occurred elsewhere. The judge anchored that interpretation in the language of section 17 (defining copying) and in authority about ephemeral copies (e.g. RAM chips), concluding that the “making” of an article must be a discrete act referable to the resulting record (for example, something downloadable), and that the mere process of training which produced weights that encode learned patterns was not itself a statutory “making” of an infringing copy within the UK unless the material intangible record had been produced in the jurisdiction. The judgment therefore resisted an expansive territorial reach that would treat every model that was the product of overseas training as an infringing article within the UK.
Practical effect: The English decision is (a) more restrictive on the question of extraterritorial primary infringement and (b) appears to relies heavily on a purposive reading of the CDPA namely that unless the “article” (i.e. the copy or its equivalent) is made in the UK, the statutory provisions directed at distribution/import cannot be invoked to reach models that were trained abroad, absent independent facts showing a UK based making or distribution of infringing material.
No Reconcilliation
The two approaches reflect different methodological choices:
- Germany: functional (technology-neutral) reading of reproduction, emphasis on effect (memorisation & local availability of outputs) and judicial fact finding using direct comparison between original and output in the jurisdiction. Also, willingness to characterise parameterised storage as fixation.
- England: textual and doctrinal caution, insistence on conventional notions of “copy” as an identifiable material fixation for the statutory scheme to bite in relation to imported or “made” articles. A clear reluctance to extend primary infringement to cover the creation of model weights which, although derivative of training, do not themselves retain the protected works.
For AI companies this divergence matters: the same engineering facts, training on a global dataset hosted on U.S./cloud servers and a service made available in Europe, can produce opposite results depending on whether a court gives weight to a functional “memorisation-as-fixation” argument (Germany) or to a more textual conception of “copy” tethered to the position of making and physical/digital transfer (England).
Evidence Relied on by the Courts: Model Functioning, Expert Testimony, Experiments, and “Direct Comparison”
The two courts also differed markedly in what they treated as appropriate proof that a model had “memorised” training inputs or that outputs amounted to reproductions.
a.) Germany: Outputs and Direct Comparison as Probative Evidence (and §286 ZPO)
A key feature of the Munich judgment is its acceptance that the court may determine memorisation by comparing the original work with the model output resulting from a simple prompt. Surprisingly, this methodology, relying on the court’s ability under §286 ZPO to form a conviction on the basis of the evidence presented, was treated as sufficient to establish memorisation in the present case. The court stated that, by performing such a comparison, it could be satisfied that the work had been memorised, and therefore reproduced within the meaning of §16 UrhG. Precisely, the Munich court accepted documentary evidence consisting of screenshots of chats, prompts and the Chatbot outputs and a subsequent side-by-side comparison of those outputs with the claimant’s original lyrics. The court regarded the deterministic character of the underlying model (as to the model’s Softmax output) and the design choices concerning decoding (which may introduce randomness at the chatbot level) as important technical inputs to its decision.
In short, because the model parameters are fixed after training and deterministic at the model level, identical or near identical outputs following simple prompts can be direct proof of memorisation, unless the defendant can identify a plausible alternative explanation (such as contrived prompts or probabilistic coincidence).
b.) England: Reliance on Technical Experts, Agreed Primer and Scepticism about Isolated Experiments
The English High Court placed central emphasis on expert evidence and the Parties’ jointly produced technical documents. Both sides called eminent technical experts (Professor Hany Farid for Getty and Professor Thomas Brox for Stability). The judge recorded that the experts produced an Agreed Technical Primer at her request and that the largely consensual expert material informed the court’s understanding of how Stable Diffusion operates. The court explicitly accepted the experts’ evidence where unchallenged and relied on the technical primer to frame factual and legal analysis.
On the evidential front, the English approach exhibited a few notable features:
Careful weighing of experimental annexes: The court examined experimental evidence (annexes of “watermark experiments”, screenshots of prompt/output pairs) but did not treat isolated experiments as dispositive unless they were representative, reliably controlled, and connected to real world use. Several of Getty’s annexes (some experiments and user prompts) were regarded as contrived and unrepresentative and the court was reluctant to infer widespread memorisation from such materials unless rooted in a broader evidential basis.
Statutory construction shaped by technical findings: The court’s statutory conclusions about “infringing copies” were informed by its technical understanding that inference, as an act, does not require training data and that the model in its final form does not contain literal stored copies of training images. Rather, it encodes patterns in weights. That technical view fed into the judge’s conclusion that the model weights were not themselves an infringing copy under the CDPA when produced abroad.
c.) Comparative Observation: Differing Standards of Proof and Different Roles for “Experiments”
Germany: gives the court a more interventionist role (direct comparison between original and output, with §286 ZPO permitting judicial factual conviction on the basis of such documentary comparisons). Experimental outputs produced in controlled litigation contexts (e.g. the claimant prompting the model) were sufficient to prove memorisation as long as the court could attribute the output to the defendant’s model and exclude convincing alternative explanations.
England: shows a stronger preference for technical expert synthesis, an insistence on representativeness and real world evidence, and doctrinal caution in inferring systemic memorisation from a small set of contrived prompts or isolated outputs. The English court explicitly credited agreed expert material and required more than isolated forensic experiments to conclude systemic memorisation leading to infringement in the forum.
How the Courts Justified their Respective Conclusions
The Munich court’s reasoning is anchored in several doctrinal moves:
- Technology neutrality of reproduction: The court emphasised that the concept of “reproduction” in §16 UrhG is technology neutral and should not be confined to classical file based or contiguous storage. A fixation that enables sensory perception (even indirectly via parameterised storage and later regeneration) qualifies as a reproduction. This understanding underpins the court’s willingness to treat parameter decomposition as a material fixation.
- Training-data creation vs model training under §44b UrhG: The Munich court read §44b UrhG (which provides for certain exceptions for text and data mining) narrowly. It covers reproductions necessary for creating the training dataset, but not reproductions that take place during the training of the model itself, because training has a purpose beyond mere text and data mining. That carve out meant the court did not treat the training process as automatically shielded by the text and data mining exception.
- Operator liability via control (“Tatherrschaft”): The court held that operators of the model exercise sufficient control (Tatherrschaft) over the model to render them primary infringers for outputs, unless the user’s prompt amounts to a loss of that control, which the court suggested could happen only with highly specific user provocation, because simple prompts do not strip “Tatherrschaft”. This finding supports the imposition of injunctive relief and primary liability on the operator.
- Procedurally pragmatic evidential standard: The court adopted a pragmatic evidential methodology (§286 ZPO) that permits judges to form convictions based on documentary comparison and technical explanation. That pragmatic evidential posture allows courts to rely on comparative outputs as decisive evidence.
- Policy overlay: Munich’s approach emphasises robust protection for rightholders and a “forward looking doctrinal adaptation” to new technical modalities of storage and reproduction.
The English court’s reasoning is anchored in textualism, careful statutory construction and the centrality of acceptable proof:
- Textual statutory analysis: The English court carefully read the CDPA definitions (particularly section 17 and the meaning of “copy”, “article” and the interplay with sections 22, 23 and 27). That reading led to the conclusion that an article must amount to a copy in the sense of a produced record whose making can be located. The abstract fact that training involved transient reproductions does not, of itself, make the produced model an infringing copy.
- Reliance on accepted technical consensus: The judge relied heavily on the technical consensus embodied in the experts’ agreed primer and joint statements. Where experts agreed, the court adopted that as fact and where disagreements were narrow, the court resolved them in light of the evidence and cross examination. The approach emphasised specialised expert proof as the proper vehicle for resolving highly technical factual disputes.
- Evidential caution on outputs: The court required proof that outputs of concern were generated by real users in the jurisdiction. Contrived experiments were scrutinised and, if unrepresentative, given limited weight. This evidentiary standard influenced the court’s willingness to make systemic findings about memorisation and about the model’s likely behaviour in ordinary user scenarios.
- Narrow construction of secondary liability: The court read secondary infringement provisions (section 27 and related provisions) restrictively. It rejected Getty’s submission that the mere creation of model weights renders the model itself an infringing copy capable of giving rise to secondary liabilities under UK law. The judge preferred an interpretation that focused on tangible or clearly identifiable intangible “articles” which themselves contain the copies.
- Policy overlay: The English judgment reflects a cautious balancing impulse of protecting rights holders, but avoid imposing sweeping liabilities on AI developers in circumstances where statutory language and the evidence do not support such an outcome.
Points of Tension and Legal Consequences
a.) Is parameterised storage a “copy”?
Munich: Yes, parameters that enable later regeneration equal fixation/reproduction. The court treats the entire model as a locus of reproduction.
England: No (at least, not automatically) model weights do not, by themselves, constitute an infringing copy within the CDPA in their final iteration. The act of training (even if it required temporary reproductions) occurring abroad does not convert the resulting model into an infringing article made in the UK.
Consequences: Operators shipping services into Germany face real risk of injunctions and secondary orders based on the Munich approach, even where training occurred overseas. In England, the more restrictive approach makes primary/secondary claims premised on overseas training harder to sustain.
b.) Who is responsible for outputs?
Munich: The model operator, because of Tatherrschaft. Liability can be primary (injunctive and compensatory remedies obtainable).
England: The court recognised the operator’s role but treated several access modes differently (downloadable models vs API access vs DreamStudio hosted inference). The English judgment framed responsibility in nuanced terms and in several places distinguished between user initiated outputs from downloaded models and hosted inference where the defendant exercises greater control. The judge declined to equate mere provision of technical infrastructure with the same degree of responsibility as would justify broad primary liability in the absence of a statutory basis.
Consequences: The Munich judgment will encourage claimants to press operator liability in circumstances where the service is made available in Germany. In England, operators who limit hosted inference or avoid downloads may face lower risk of primary copyright liability, although other claims such as trade mark, passing off, or targeted secondary claims remain possible.
c.) Evidential routes to proving “memorisation”?
Munich: direct output comparison from simple prompts. Judge’s conviction formed under civil evidential rules.
England: expert synthesis, agreed technical primer, representative evidence. Contrived experiments alone are weak.
Consequences: Claimants seeking relief in Germany may rely on well crafted prompt/output pairings and direct comparative evidence. In England claimants should plan for rigorous expert proof and, ideally, evidence demonstrating occurrence in real user contexts.
Concluding Reflections: Divergence, Convergence, and Unresolved Questions
The two judgments supply complementary but conflicting doctrinal signals:
- Convergence: Both courts recognise the technical realities of contemporary generative models, their architecture, training on large scraped datasets (LAION etc. in the English case), and the phenomenon of memorisation/overfitting as a potential factual phenomenon that can cause reproductions.
- Divergence: They diverge, however, on legal characterisation and territorial reach. Germany treats parameterised storage and local availability of reproducing outputs as sufficient to ground reproduction and operator liability. England adopts a more textualist, evidence intensive and territorially limited approach.
Open issues that litigation and legislatures will need to address:
- Reconciling parameterisation with statutory concepts of “copy”: Is parameterised encoding of patterns a “fixation” for the purposes of copyright law? Munich answers affirmatively (subject to its factual findings). England is more cautious. The legal community and legislatures may need to clarify whether existing copyright categories suffice to capture modern machine learned artifacts.
- Evidential standards for “memorisation”: Will civil-law courts (with their willingness to form convictions under different procedural rules) generally be more receptive to direct output comparison evidence than common-law courts that emphasise representativeness and expert consensus? The two cases suggest the answer may be yes, at least in the near term.
The immediate takeaway is that the legal risk associated with AI model training and deployment is heavily jurisdiction dependent and that litigation strategy must be informed by careful technical forensics, robust expert instruction, and early consideration of forum advantages.
