In Getty Images v Stability AI [2025] EWHC 2863 (Ch), handed down on 4 November 2025, the England and Wales High Court (Chancery Division) found that model weights do not store copies of training images in any legally recognisable sense, causing the secondary copyright infringement claim to fail, while only limited trade mark findings were made in Getty's favour. The judgment does not end AI copyright litigation in the United Kingdom; it clarifies one narrow technical question while leaving the broader questions of direct infringement during training, the scope of the text and data mining exception, and the liability of AI developers for outputs largely open. For international companies building or deploying generative AI models with any connection to the UK, the decision demands careful re-reading rather than relief.
The Judgment in Context: What the Court Actually Decided
The proceedings concerned Getty Images' claim that Stability AI had, in training its Stable Diffusion model on a large corpus that included images licensed through Getty's platforms, committed acts of copyright infringement and trade mark infringement under UK law. The Chancery Division's November 2025 ruling addressed, among other issues, the technical architecture of the trained model and whether the resulting weights could be characterised as unauthorised copies of the claimant's works.
The court held that model weights do not store copies of training data in the sense required to establish the copying right under the Copyright, Designs and Patents Act 1988. This finding disposed of the secondary infringement claim, which had alleged that distributing or using the trained model was equivalent to distributing infringing copies. The reasoning turned on an analysis of how floating-point numerical parameters encode statistical relationships rather than preserve pixel-level reproductions of source material.
The trade mark findings were described in the judgment as limited. The court acknowledged that certain outputs generated by the model could, in specific circumstances, reproduce watermark-style elements associated with Getty's registered marks, but the scope of liability arising from that finding was narrow and subject to further argument. No finding of wide-ranging trade mark infringement was made at this stage.
Critically, the judgment did not rule on whether the act of ingesting and processing copyright-protected images during the training phase itself constitutes primary infringement of the reproduction right. That question, and the related question of whether the UK's text and data mining (TDM) exception under section 29A of the CDPA applies to commercial AI training, remains unresolved on the facts of this case.
The Text and Data Mining Exception: A Persistent Fault Line
Section 29A of the CDPA, introduced in 2014 and amended incrementally since, permits copying for the purpose of computational analysis where the person undertaking that analysis has lawful access to the work. The UK Government's position, articulated in policy consultations preceding the judgment, has been to consider whether to expand TDM permissions for commercial AI developers – a proposal that attracted significant opposition from rights holders, including the creative industries bodies that supported Getty's litigation strategy.
The court in Getty v Stability AI did not need to decide the TDM point definitively, because the secondary infringement claim failed on the weights analysis. This means that the commercial TDM question – whether a generative AI company training on internet-scraped data without a licence can rely on the exception – will need to be litigated or legislated before the UK offers genuine legal certainty. Developers who assumed the November 2025 ruling gave them a green light on training are reading it too broadly.
The contrast with developments in Germany is instructive. In GEMA v OpenAI (LG München I, 42 O 14139/24, 11 November 2025), the court found that memorisation of lyrics by GPT-4 and GPT-4o infringed the relevant provisions of the UrhG, with the German TDM exception under paragraph 44b found inapplicable on the facts. OpenAI is appealing that decision. A further LG München I ruling, reported on 31 July 2026 and described as not yet final, indicated that training on protected music may infringe even where the training itself took place outside the EU – a proposition with significant implications for any developer processing EU-resident rights holders' works on non-EU infrastructure. That ruling should be treated as non-final and subject to appeal, but its reasoning merits close attention.
Taken together, the UK and German proceedings illustrate that no single jurisdiction has yet produced a stable, comprehensive answer to the training-phase copyright question. Rights holders are forum-shopping across the EU and the UK simultaneously, and developers must map their exposure in each relevant territory.
If your organisation is assessing AI training compliance obligations across multiple European jurisdictions, our team would be glad to discuss the applicable frameworks. Please write to info@ferrazwhitmore.com or visit our contacts page.
Design Rights and the Overlooked Visual Dimension
Beyond copyright infringement, the proceedings touched on design registration and the protections available to image owners under the UK Registered Designs Act 1949 and the retained EU law framework that continues to apply in Great Britain. While the judgment's primary focus was copyright, the underlying dispute highlights a broader vulnerability for AI developers: the visual outputs of generative image models may reproduce protected design elements, not merely artistic works protected by copyright.
Design registration protects the appearance of a product or part of a product, including its lines, contours, colours, shape, and ornamentation. Where a generative model is prompted to produce images in a style that closely replicates registered design features – whether in fashion, product packaging, or graphic design – developers may face parallel claims under design law even if they successfully defeat copyright infringement claims on the weights theory. The November 2025 judgment does not provide a defence to those claims; it simply did not address them on the facts before the court.
Trade Mark Exposure: Limited Findings, Broader Lessons
The limited trade mark findings in the judgment nevertheless carry practical weight. The court's acceptance that certain model outputs could reproduce elements associated with a registered trade mark confirms that trade mark law is a live risk vector for generative AI, not merely a theoretical one. For developers whose models are capable of generating branded imagery, watermarks, logos, or product photographs, the risk of trade mark infringement – separate from and in addition to copyright infringement – requires independent analysis.
UK trade mark law, governed by the Trade Marks Act 1994 and the retained framework from the EU Trade Marks Directive, protects signs registered with the Intellectual Property Office. Infringement can occur through use of a sign identical or similar to a registered mark in the course of trade, without the need to demonstrate copying in the copyright sense. An AI model that consistently generates outputs bearing a watermark or logo associated with a registered mark may, depending on how those outputs are used commercially, cross the threshold of infringing use.
Procedural and Corporate Considerations for Non-UK Defendants
Stability AI is incorporated in England and Wales, meaning service of proceedings under the Civil Procedure Rules (CPR) was straightforward and its UK presence was confirmed through Companies House registration. For foreign AI developers without a UK subsidiary, the procedural position is more complex. Service out of the jurisdiction under CPR Part 6 requires the claimant to establish a good arguable case on a jurisdictional gateway – for example, that the defendant committed acts within the jurisdiction, or that a contract was concluded or breached in England and Wales.
The practical implication is that developers structured entirely outside the UK, with no UK entity on the Companies House register, are not automatically immune from proceedings. If a model is made available to UK users, if training data was scraped from UK-based servers, or if the developer has any business presence in Great Britain, claimants will seek to invoke jurisdiction. The November 2025 judgment, by establishing that proceedings against a UK-registered AI developer can succeed at least partially, increases the incentive for rights holders to pursue foreign developers through service-out applications.
Regulatory considerations beyond IP also arise. Where an AI developer's products touch on financial services or regulated data, FCA authorisation requirements and OFSI sanctions compliance must be considered separately from the IP litigation risk. These are not addressed by the November 2025 judgment but form part of the overall compliance matrix for any company operating AI systems in or from the United Kingdom.
For detailed guidance on IP litigation strategy and procedural options before the UK courts, our litigation and arbitration practice can assist: see our UK litigation and arbitration services page for an overview of our approach.
Cross-Border Implications: EU AI Act and Pending CJEU Guidance
The judgment arrives at a moment when the European Union is developing its own regulatory architecture for AI, and when the CJEU is preparing to provide guidance on related questions of platform liability for intellectual property. The case of CJEU C-250/25 Like Company v Google had its hearing on 10 March 2026, with the Advocate General's opinion expected in September 2026. Although that case concerns different facts, the AG opinion will address principles of online platform liability for IP infringement that may inform how EU courts approach the liability of AI developers whose outputs infringe third-party rights.
Our earlier analysis of Like Company v Google discusses the broader implications of that referral for platform operators: see our deep-dive on CJEU C-250/25.
For clients operating across both the UK and EU, the divergence in applicable law is now substantial. Post-Brexit, the UK is not bound by the EU AI Act or by CJEU jurisprudence interpreting the InfoSoc Directive's TDM provisions. UK courts will develop their own doctrine, influenced but not constrained by EU developments. This creates a genuine bifurcation risk: a training practice that is permissible in the UK may infringe EU law, and vice versa, depending on how the relevant exceptions and the territorial scope of infringement are interpreted in each jurisdiction.
Clients seeking a comparative overview of AI compliance obligations under the EU AI Act framework, including obligations affecting developers based in non-EU countries whose systems are deployed in the EU, may find it useful to consult our AI Act compliance services overview, which addresses the practical steps required of both EU and third-country operators.
Strategic Recommendations for AI Developers
The November 2025 judgment reduces one category of risk – secondary infringement through the distribution of trained weights – but does not eliminate the IP exposure of generative AI developers in the UK. The following considerations are relevant to any organisation reviewing its position in light of the ruling:
- Audit training data provenance. The weights-do-not-copy finding applies to the trained model, not to the acts performed during training. Developers should document the source, licensing status, and territorial origin of training data, distinguishing between content that was publicly scraped, content obtained under licence, and content generated or commissioned by the developer.
- Review TDM reliance carefully. Reliance on the UK TDM exception requires lawful access to the work and is subject to conditions that have not been comprehensively tested in commercial AI contexts. Do not assume the exception applies without analysis of access conditions and the commercial nature of the use.
- Map output risks separately from training risks. The trade mark findings in the November 2025 judgment confirm that output-level claims remain viable. Developers should test models for their propensity to reproduce registered marks and implement filtering or training controls where necessary.
- Consider corporate structure and jurisdictional exposure. The presence or absence of a UK entity affects service of proceedings and enforcement, but does not eliminate substantive liability where UK-resident rights holders' works are involved.
- Monitor legislative developments. The UK Government's ongoing review of AI and IP policy may result in legislative changes that alter the TDM framework. Compliance strategies should be designed with this uncertainty in mind and reviewed as legislation evolves.
- Track the German proceedings. The LG München I decisions in GEMA v OpenAI and the subsequent non-final ruling on extraterritorial training are on appeal. Their outcome will affect the position of any developer whose training infrastructure is located outside the EU but whose systems interact with EU-resident works or users.
Outlook: Litigation Risk Remains Elevated
The November 2025 ruling is a significant data point, not a resolution. The High Court's analysis of model architecture will be studied carefully, but the technical understanding of how neural networks encode information continues to develop, and future claimants may advance different expert evidence or target different aspects of the training and deployment pipeline. The primary infringement question – whether ingestion of copyright works during training itself infringes the reproduction right – will require its own litigation or authoritative legislative guidance before UK law can be considered settled.
Rights holders have demonstrated a sustained appetite for IP litigation against AI developers across multiple jurisdictions. The partial success in the November 2025 proceedings – particularly the trade mark findings and the court's willingness to engage with technical evidence about model behaviour – confirms that these cases are litigable and that courts are prepared to award meaningful findings, even if not the sweeping injunctive relief some claimants sought initially.
For AI developers, the appropriate response is neither complacency nor paralysis, but careful, jurisdiction-specific legal advice that takes account of the current state of the law and the direction of travel in the courts and legislatures of each market where the developer operates.
If you would like to discuss how the Getty v Stability AI judgment affects your organisation's IP position in the UK or in other jurisdictions where Ferraz & Whitmore operates, please contact us at info@ferrazwhitmore.com or through our contacts page. We can assist with risk mapping, licensing strategy, and representation in proceedings before the UK courts and international arbitral tribunals.
Disclaimer: This publication is provided for informational purposes only and does not constitute legal advice. The information herein should not be relied upon as a substitute for professional legal counsel tailored to your specific circumstances. Ferraz & Whitmore assumes no liability for actions taken or not taken based on the contents of this material. For advice regarding your particular situation, please contact info@ferrazwhitmore.com.