Authored By: Malachi Rees-Morny
University of Exeter
In November 2025, the High Court delivered a judgment in Getty Images v Stability AI, which may very well be one of the most significant intellectual property decisions of the modern AI era. The claim stems from Stability AI’s alleged use of millions of Getty’s professionally licensed images to train its Stable Diffusion AI model without authorisation or remuneration. Litigation of this magnitude was required to test the boundaries of the existing framework, and that in itself is a diagnostic: a statute designed for human creators that operate in analogue markets was being pressed into service to govern large-scale machine learning (ML) systems whose commercial value derives, in large part, from mass-scale appropriation of copyright-protected content.
Whether the United Kingdom’s copyright law provides a satisfactory economic rights framework for generative AI (GenAI) requires analysis of two distinct phases of the AI lifecycle: The ‘input’ phase concerns the use of protected works to train AI models; the ‘output’ phase concerns the ownership and protectability of the content those models produce. This article argues that the Copyright, Designs and Patents Act 1988 (CDPA) offers a principled but structurally incomplete economic framework. The CDPA successfully resists the subsidy to AI development that a broad text and data mining (TDM) exception would represent, and the Getty judgment modernises secondary infringement doctrine in important respects, but territorial weakness, uncertain subsistence, and the absence of a structured licensing mechanism leave both rightsholders and developers in a state of chronic legal uncertainty. The article covers the relevant legal framework, analysis of the Getty judgment, critical evaluation of the framework’s adequacy for each stakeholder group and finally the comparative perspectives. The conclusion will then propose a possible path to reform.
The Legal Framework
Issue I – The Input Problem: The Scope of Section 29A
The treatment of reproduction during the training of these ML models is the economic core of the GenAI conflict. Section 17 of the CDPA states that ‘copying’ a work includes storing it in any medium by electronic means.[1] The consumption of millions of images into a training pipeline satisfies this definition, so the absence of a license for such copying constitutes prima facie infringement unless an exception applies.
The UK’s TDM exception, codified in section 29A of the CDPA, permits copying only in cases of ‘non-commercial research.’[2] Its scope is substantially narrower than the regime established by Article 4 of the EU Digital Single Market Directive, which permits commercial TDM subject to a rightsholder opt-out.[3] Proposals to introduce a broad commercial TDM exception were advanced by the UK Government in 2022 but were subsequently abandoned following sustained opposition from the creative industries. The December 2024 consultation saw an opt-out model analogous to the EU position, but the March 2026 Government Report ultimately rejected this approach, concluding that the technical difficulty of implementing effective opt-outs and the risk to the creative sector’s gross value added outweighed the anticipated benefits to AI development.[4] Consequentially, commercial AI training on copyright-protected material without a license constitutes infringement under English law wherever copying is present within the jurisdiction.
From a purely economic standpoint, the current position corresponds to the theoretical framework proposed by Landes and Posner, who argue that copyright must provide an optimal incentive to invest in creative production.[5] By preserving the right to reproduction against commercial exploitation, the CDPA sustains the licensing revenues upon which the creative industries depend. Critics drawing on the Bruegel working paper tradition respond that mandatory individual licensing for AI training inputs generates prohibitive transaction costs, entrenching incumbents capable of bearing those costs and suppressing the productivity gains that AI could otherwise deliver.[6]
Notably, the March 2026 Government Report proposed that s.9(3) be removed entirely, on the basis that it is unclear, is hardly relied upon in practice, and is omitted from the copyright law of any comparable jurisdiction.[7] Should this proposal be enacted, the tension between s.9(3)’s ownership allocation and the originality threshold present in Infopaq would be resolved by eliminating s.9(3) rather than clarifying both propositions, likely leaving AI-generated works without any specific statutory provision.
Issue II – The Output Problem: Section 9(3) and Originality
The second dimension of the framework concerns the ownership and protectability of AI-generated works. The UK is unusual internationally in providing an express statutory provision: section 9(3) of the CDPA deems the ‘author’ of a computer-generated work to be ‘the person by whom the arrangements necessary for the creation of the work are undertaken.’[8] This provision resolves the ownership question but does not address the antecedent issue of whether such a work is protectable at all.
Copyright subsists in an artistic or literary work only if it is ‘original.’ Following the Court of Appeal’s treatment of Infopaq in THJ Systems Ltd v Sheridan, the originality standard applicable in UK law requires the work to reflect the ‘author’s own intellectual creation,’ constituted by the author’s free and creative choices.[9] This standard sits in considerable tension with the automated nature of generative output. Where a model produces an image or text in response to a brief prompt, it is difficult to identify the human intellectual choices that would satisfy the Infopaq threshold in any substantive sense. The House of Lords Communications and Digital Committee’s March 2026 Report emphasised that speed and volume of AI output are not substitutes for the value of human creativity, signalling a strict approach to subsistence.[10] Through Sögüt Atilla’s analysis of s.9(3), exposing how: the person who makes the necessary arrangements owns the work under section 9(3), but those arrangements may not generate the originality that copyright requires to subsist; something of a subsistence limbo can be observed in the UK’s operations.[11]
Getty Images v Stability AI [2025]: The Main Event
Issue III – Territoriality and the Failure of Primary Claims
Getty Images (US) Inc v Stability AI Ltd was the first English case to directly confront the copyright implications of training a large-scale image-generation model on commercially owned photographic archives.[12] By the time of the trial, the primary infringement claims were dropped by Getty (not dismissed by the court on territorial grounds) as the litigation progressively narrowed in scope. English courts have thus still not resolved whether training GenAI models on copyrighted works constitutes primary infringement under s.17 CDPA. The territorial exposure that commenters have identified is that AI developers may offshore the training process to jurisdictions with permissive TDM regimes like Japan’s broad research-neutral exception, while deploying the resulting model commercially within the UK. A superficial framework, fundamentally incapable of capturing extra-territorial activity. This remains a vulnerability inherent in the UK’s model and fails to provide practical economic protection.
Issue IV – The Secondary Infringement Breakthrough
The transformative aspect of the judgment lay in the Court’s analysis of secondary infringement under sections 22 and 23 of the CDPA.[13] Getty argued that Stability AI had imported and dealt in an ‘article’ constituting an ‘infringing copy’ in the form of the model weights. The Court made two consequential findings of principle. First, it held that an ‘article’ within the meaning of the CDPA may be intangible, extending the concept to encompass the digital weights of an AI model. This represents a material departure from the tangibility assumptions embedded in EU law and a significant modernisation of English copyright doctrine. Second, the Court held that model weights could in principle constitute ‘infringing copies’ where generated through the unauthorised reproduction of protected works, notwithstanding that the original images were not stored in any human-recognisable form within those weights. The theory of liability turned on the derivation of the model rather than on literal reproduction.
The claim ultimately failed because the specific model weights at issue were found not to ‘contain’ copies of Getty’s works in the requisite sense. The precedent on intangible articles nevertheless significantly expands the risk profile for AI development within the jurisdiction. As Kretschmer and others observe, the lifecycle of machine learning models creates points of potential liability that did not previously feature in standard IP risk assessments for technology companies.[14]
Issue V – Trade Marks and the Watermark Problem
The Court further found limited trade mark infringement under section 10(2) of the Trade Marks Act 1994, arising from the appearance of watermark-like features in certain AI-generated images bearing a confusing resemblance to Getty’s registered marks.[15] This element of the judgment illustrates that, where copyright provides an incomplete remedy, trade mark law and the tort of passing off offer supplementary protection for brand identity and commercial reputation. The Court was careful to confine these findings to the specific facts, resisting any suggestion that AI-produced imitation of visual style attracts liability as a general matter.
Critical Evaluation
Issue VI – From the Perspective of the Creative Industries
The framework is increasingly satisfactory in principle but failing in enforcement. The rejection of a broad commercial TDM exception preserves the reproduction right as the economic foundation of the creative industries’ licensing model. The House of Lords Report correctly identifies this as essential to maintaining the ‘gold standard’ of UK IP protection.[16] The most difficult issue is the territorial limitations exposed by Getty to permit systematic circumvention of those rights without legal consequence, provided training occurs abroad. The absence of mandatory transparency obligations exacerbates this issue, as without disclosure of the datasets used in model training, rightsholders have trouble identifying the infringement necessary to found a claim.
Issue VII – From the Perspective of AI Innovation
The framework is arguably unsatisfactory for AI developers operating in or serving the UK market. The legal uncertainty surrounding section 9(3), the potential secondary liability arising from the Getty precedent on intangible articles, and the absence of a safe harbour for commercial TDM collectively create a chilling effect on domestic AI investment. The transaction costs of negotiating individual data licenses across a dispersed landscape of rightsholders, each able to set arbitrary prices without reference to any objective benchmark, represent precisely the market failure that critics of strong copyright protection have identified as the principal economic cost of an overly protective regime in the AI context.
Issue VIII – The Licensing Horizon
There is a growing consensus that a market-led licensing economy represents the only durable resolution of these competing interests. Rosati argues that AI training is not fully covered by any existing copyright exception, making an effective licensing obligation inescapable, and that the appropriate response is to build the infrastructure that would make that obligation workable.[17] The March 2026 Government Report’s rejection of a broad opt-out in favour of market-led solutions implies that AI developers will be expected to pay for training data access. If supported by collective licensing infrastructure, this approach could rebalance the economic relationship between developers and creators while providing the transactional certainty that investment requires.
IX – Comparative Perspectives
The UK’s position sits between two polar alternatives. The EU retains a commercial TDM exception subject to the DSM Directive’s rightsholder opt-out, while AI-specific transparency obligations under the AI Act create an emerging framework for accountability. This regime offers greater certainty to both sides but has been criticised for underprotecting creators in practice, given the difficulty of implementing effective opt-outs at scale. Japan has adopted the opposite approach, exempting all forms of data mining from copyright liability regardless of commercial purpose, treating training data access as essential economic infrastructure.
The Beijing Internet Court’s 2023 decision in Li v Liu offers a contrasting perspective on the output question.[18] The court found that a user’s selection of prompts, negative prompts, and generation parameters reflected sufficient personal style to satisfy Chinese originality requirements. This conclusion appears materially inconsistent with the cautious approach signalled by the House of Lords Committee in the UK context. This deviation shows the absence of any emerging international consensus on the appropriate protection threshold for AI-generated works and creates conditions for regulatory arbitrage and forum-shopping that a purely domestic solution cannot address.
Conclusion
The UK copyright framework, as anchored by the CDPA 1988 and interpreted in Getty Images v Stability AI, provides a coherent but structurally incomplete economic framework for generative AI. The preservation of the reproduction right against commercial exploitation maintains the economic incentives on which creative production depends, and the Getty judgment’s recognition of intangible articles as potential infringing copies represents genuine doctrinal progress. However, territorial vulnerability permits systematic circumvention of those rights, subsistence uncertainty under the Infopaq standard leaves AI-generated works in an indeterminate condition, and the absence of transparency and collective licensing obligations imposes heavy enforcement costs on rightsholders while generating legal uncertainty for developers.
A satisfactory framework requires proactive legislative intervention on three fronts. Parliament should enact mandatory transparency obligations requiring AI developers to disclose training datasets, enabling rightsholders to identify and pursue infringements. A statutory collective licensing regime would reduce transaction costs while preserving creators’ economic entitlements. Finally, the originality threshold applicable to AI-assisted works should be clarified by statute to provide the certainty that incremental judicial development cannot deliver at adequate speed. Until such reforms are implemented, the UK risks failing both groups its copyright framework is designed to serve: failing creators by permitting extra-territorial circumvention of their rights, and failing developers by maintaining a landscape too uncertain to sustain domestic AI investment at the scale the sector demands.
Bibliography
Primary Sources
Legislation
Copyright, Designs and Patents Act 1988
Directive (EU) 2019/790 of the European Parliament and of the Council of 17 April 2019 on copyright and related rights in the Digital Single Market [2019] OJ L130/92
Trade Marks Act 1994
Cases
Getty Images (US) Inc v Stability AI Ltd [2025] EWHC 2863 (Ch)
Infopaq International A/S v Danske Dagblades Forening (C-5/08) EU:C:2009:465
Li v Liu (Beijing Internet Court, 2023)
THJ Systems Ltd v Sheridan [2023] EWCA Civ 1354
Secondary Sources
Atilla S, ‘Dealing with AI-Generated Works: Lessons from the CDPA Section 9(3)’ [2024] 19(1) JIPLP 43 <https://doi.org/10.1093/jiplp/jpad102\>
Communications and Digital Committee, AI, Copyright and the Creative Industries (4th Report of Session 2024–26, HL Paper 267, 2026)
Kretschmer M and others, ‘Copyright Law, and the Lifecycle of Machine Learning Models’ (2024) 55 IIC, 110
Landes W and Posner R, ‘An Economic Analysis of Copyright Law’ (1989) 18 J Legal Stud 325
Martens B, ‘Economic Arguments in Favour of Reducing Copyright Protection for Generative AI Inputs and Outputs’ (Bruegel Working Paper No 09/2024, 2024)
Rosati E, ‘Copyright Exceptions and Fair Use Defences for AI Training Done for “Research” and “Learning,” or the Inescapable Licensing Horizon’ (2025) 16(3) EJRR
UK Government, ‘Report on Copyright and Artificial Intelligence’ (March 2026)
[1] Copyright, Designs and Patents Act 1988, s 17.
[2] Ibid, s 29A.
[3] Directive (EU) 2019/790 of the European Parliament and of the Council of 17 April 2019 on copyright and related rights in the Digital Single Market [2019] OJ L130/92 (‘DSM Directive’), art 4.
[4] UK Government, ‘Report on Copyright and Artificial Intelligence’ (March 2026).
[5] William Landes and Richard Posner, ‘An Economic Analysis of Copyright Law’ (1989) 18 J Legal Stud 325, 326.
[6] Bertin Martens, ‘Economic Arguments in Favour of Reducing Copyright Protection for Generative AI Inputs and Outputs’ (Bruegel Working Paper No 09/2024, 2024), 3.
[7] UK Government, ‘Report on Copyright and Artificial Intelligence’ (March 2026).
[8] Copyright, Designs and Patents Act 1988, s 9(3).
[9] Infopaq International A/S v Danske Dagblades Forening (C-5/08) EU:C:2009:465; THJ Systems Ltd v Sheridan [2023] EWCA Civ 1354.
[10] Communications and Digital Committee, AI, Copyright and the Creative Industries (4th Report of Session 2024–26, HL Paper 267, 2026), 3-5.
[11] Sögüt Atilla, ‘Dealing with AI-Generated Works: Lessons from the CDPA Section 9(3)’ [2024] 19(1) JIPLP 43.
[12] Getty Images (US) Inc v Stability AI Ltd [2025] EWHC 2863 (Ch), para 525.
[13] Copyright, Designs and Patents Act 1988, ss 22–23.
[14] Martin Kretschmer and others, ‘Copyright Law, and the Lifecycle of Machine Learning Models’ (2024) 55 IIC, 110-138.
[15] Trade Marks Act 1994, s 10(2).
[16] Communications and Digital Committee (n 10), 46-47.
[17] Eleonora Rosati, ‘Copyright Exceptions and Fair Use Defences for AI Training Done for “Research” and “Learning,” or the Inescapable Licensing Horizon’ (2025) 16(3) EJRR, 979.
[18] Li v Liu (Beijing Internet Court, 2023).
