The last battle — AI, copyright and the shadow of Donaldson

Following the landmark judgment in Getty Images v Stability AI, Dr Corsino San Miguel considers questions of copyright, authorship and the governance of knowledge in the age of machine learning.

The Battle of the Booksellers

The last great battle between England and Scotland was not fought with muskets at Culloden, but with pamphlets, lawsuits and reprinted verse. It did not take place on a field of heather, but across courtroom benches and bookseller ledgers. This was the Battle of the Booksellers: a three-decade war of ideas and commerce in which Scottish printers challenged the monopolies of London’s publishing elite. At its heart lay a question that still echoes today: who owns knowledge once it has been shared?

In 1774, that long war of ledgers and letters found its end in the House of Lords. In Donaldson v Beckett, the Lords declared that copyright was not a perpetual birthright but a finite privilege – a creature of statute, not of common law. The statute in question, the Statute of Anne (1710), had introduced a clear limit: 14 years, renewable once if the author lived. Alexander Donaldson, an Edinburgh printer, had defied London publishers by reprinting works whose statutory protections had expired. The Lords upheld his right to do so, dismantling the notion that publishers held eternal dominion over the printed word. It was a ruling not for commerce but for the public – an affirmation that the circulation of knowledge serves society more than it serves the ledger.

Two-and-a-half centuries later, the battle returns in digital guise. The courtroom endures; the battlefield has moved. Today, not books but training data lies at the heart of the conflict. In the landmark judgment handed down on 4 November 2025, Getty Images v Stability AI, a new class of publishers – those who train machines rather than print pages – stands accused of overreach. The 18th-century tensions between exclusivity and access, between rightsholders and re-users, persist – now amplified by global networks and machine learning at scale.

The echoes of Donaldson v Beckett reach further than we think. The Lords of 1774 faced not merely a quarrel over profits but a constitutional question: who governs knowledge? Their answer was unambiguous. It is Parliament, not the courts, that defines the limits of property in ideas. Copyright, they insisted, is a tool of policy, not a moral entitlement; a mechanism of encouragement, not an instrument of monopoly.

Two-and-a-half centuries on, the Getty litigation shows how that settlement is fraying. Once again, judges were asked to decide whether a new technology – artificial intelligence (AI) – trespasses upon the boundaries of creative ownership. And once again, the judiciary declined to legislate by analogy. The High Court’s ruling in Getty left the core question unanswered, not from hesitation but from constitutional restraint: the Copyright, Designs and Patents Act 1988 (CDPA) simply does not imagine machines learning from the world’s collective memory.

Where Donaldson affirmed that no perpetual monopoly can stand without statutory warrant, Getty affirms that no digital monopoly can be condemned without one. Whether model training amounts to copying or adaptation, whether neural weights are ‘articles’, or whether data-mining falls within fair dealing – these are not questions of doctrine but of legislative design.

The judgment: a victory deferred

The decision in Getty Images v Stability AI Ltd is less a verdict than a mirror held up to the limits of contemporary copyright law. Hailed as the first full-scale trial to test how the CDPA meets machine learning, it revealed that the statute – like the Statute of Anne three centuries earlier – was not written for the technology it now faces. Getty alleged that Stability appropriated millions of its photographs, watermarked and catalogued on Getty’s sites, as unlicensed training data for Stable Diffusion, a generative-image model capable of producing pictures from text prompts.

The narrowing of claims

At the outset, Getty pleaded primary and secondary copyright infringement, database-right infringement, trademark infringement and passing off. As the evidence unfolded, that edifice collapsed under jurisdictional reality. Cross-examination and disclosure indicated that the relevant development and training activities occurred outside the UK on cloud infrastructure. Under orthodox territoriality, infringing acts must occur within the jurisdiction. No server log, no witness, tied the core copying to the UK. By June 2025, Getty withdrew its primary copyright and database-right claims, acknowledging it could not prove where the copying occurred.

This was not merely a forensic loss; it was a cartographic lesson. The CDPA’s map presumes that copying has a place. Distributed computing dissolves that premise. Where the Statute of Anne could draw a jurisdictional circle around London’s presses, our statutory lines run thin across fibre and cloud.

The secondary infringement theory

Having ceded the battlefield of training, Getty regrouped on distribution. Its remaining claim turned on secondary copyright infringement under Sections 22 and 23 of the CDPA: even if the alleged copying had taken place abroad, the importation or commercial dealing of an ‘infringing article’ in the UK could still amount to infringement.

Getty’s case rested on an ambitious reading of Section 27(3): that the model files of Stable Diffusion – its trained neural-network weights in versions v1, v2, XL and 1.6 – should themselves be treated as ‘articles’ whose making would have constituted an infringement had it occurred in the UK. In other words, making those models available for download or use within the jurisdiction amounted, in Getty’s view, to importing infringing copies.

The court addressed two novel questions:
(i) whether electronic files such as model weights can be an ‘article’ for the purposes of Sections 22 and 23; and
(ii) whether such an article could be an ‘infringing copy’ under Section 27.

On the first issue, Mrs Justice Joanna Smith held that an article need not be tangible: the statutory concept could indeed extend to electronic copies stored in intangible form. On the second, however, she drew a crucial line. Getty had accepted, in line with both parties’ expert evidence, that the Stable Diffusion model weights do not contain, store or reproduce any Getty images. The judge therefore rejected the argument that it was enough that the model’s making would have infringed if done in the UK.

An ‘infringing copy’, she ruled, must at some point have consisted of, contained or stored a copy of a protected work. Because the model weights never did, Stable Diffusion could not qualify. The finding preserved the logical architecture of the CDPA: the act of copying, not the potential to copy, defines infringement. Getty’s construction, elegant in theory but unanchored in statutory language, was thus refused.

Trademarks and the ghost of the watermark

Getty also pursued trademark infringement under Sections 10(1), 10(2) and 10(3) of the Trade Marks Act 1994, alleging that versions 1, 2, XL and 1.6 of Stable Diffusion generated images bearing synthetic Getty or iStock watermarks. The evidence fell into two groups: experiments conducted by Getty’s own team using text prompts derived from Getty captions, and a few examples found ‘in the wild’.

Stability countered that there was no proof any UK user had ever produced such outputs, noting that Getty captions were not typical prompts in public use. Mrs Justice Joanna Smith agreed for the most part, finding no evidence of watermark generation by UK users for versions XL and 1.6 and dismissing those claims entirely.

For versions 1 and 2, however, the judge accepted that at least one instance of a synthetic watermark resembling the Getty or iStock mark was likely to have been generated in the UK. She identified two infringements of the iStock mark under Sections 10(1) and 10(2) and one of the Getty mark under Section 10(2). Yet she described these findings as “extremely limited”, insufficient to show any broader market impact. The Section 10(3) claim – alleging reputational harm – failed for want of evidence.

The judgment framed watermark artefacts not as a legal revolution but as a narrow, fact-specific anomaly – a product-safety issue for the developer, not a foundation for remaking trademark law.

A constitutional restraint disguised as neutrality

For some practitioners, the outcome felt anticlimactic. Constitutionally, it is notable for its restraint. The court refused to craft new categories of infringement or stretch statutory words to novel facts simply because the technology is new. It applied Donaldson in digital key: copyright is a creature of statute, not judicial invention. Where Parliament has not spoken, the judiciary must not legislate from the bench.

There is a cost to this fidelity. Stability left court largely vindicated; rightsholders left with few tools where training occurred abroad. Unless Parliament intervenes, UK claimants will struggle to police model training that happens beyond the water’s edge. The High Court has, in effect, handed the quill back to Westminster. In fact, as I discussed in a previous article,  ‘The angel’s share’, the true concern extends beyond data ownership; it lies in the gradual delegation of human reasoning to the very systems trained on it, raising profound questions about the future of creativity and authorship.

The deferred victory

If Donaldson was a victory for the public domain, Getty is a victory deferred – postponed until Parliament draws a statutory frontier around the digital realm. The judgment reflects judicial restraint rather than uncertainty: a recognition that only legislation can resolve the tension between innovation and authorship.

Following the ruling, attention has turned to Westminster. The Government’s preferred option for reform, outlined in its Copyright and AI Consultation, would make it easier for developers to use protected works for model training – a stance that has met strong resistance from across the creative industries and from several peers in the House of Lords. Concerned by what they viewed as legislative inertia, some peers sought to introduce new AI-related copyright protections into the Data (Use and Access) Bill, forcing ministers to concede to an accelerated timetable for policy development.

Working groups have since been convened to address transparency, rights control and practical licensing mechanisms, though substantive progress is unlikely before year-end. Previous attempts to secure a voluntary industry consensus have already faltered. Responsibility now rests with the government to turn consultation into policy. The High Court may have handed the quill back to Parliament, but it is Parliament that must now decide how far creativity may be automated, and under whose authority knowledge is allowed to learn itself.

Yet the aftermath of Getty also forces a constitutional reckoning. The UK proclaims parliamentary sovereignty, but its freedom to legislate on AI and copyright is increasingly shaped by transnational commitments – the Atlantic Declaration, the UK-US Memorandum of Understanding on the Technology Prosperity Deal and the G7’s shared principles on trustworthy AI. The same digital networks that enable innovation now dissolve the very boundaries of national authority, binding domestic reform to the infrastructures and alliances through which data, models and markets move. The High Court may have handed the quill back to Parliament, but the ink must now be shared – written in Westminster, yet legible in Washington, Brussels and Tokyo alike.

The last battle revisited

Getty v Stability AI is not a failure of adjudication; it is a constitutional mirror. It shows that in the age of the algorithm, the locus of authorship has shifted – and with it, the locus of governance. Just as the Statute of Anne was born of the printing press, the next settlement must arise from Parliament’s will.

The last battle for copyright was fought with pamphlets and presses; the next is fought with code, datasets and the very definition of creativity. The question is unchanged: who owns knowledge once it has been shared? If the lesson of 1774 was that perpetual rights yield to the public interest, the lesson of 2025 is that jurisdictional boundaries yield to networks.

Only Parliament can reconcile those truths. The courts have drawn the line where their mandate ends. The quill sits ready. It is for government to write.

Article written by Dr Corsino San Miguel PhD, LLB in Scots law and graduate in Spanish law; co-founded and led European Telecom Company before entering academia. He is now a member of the AI Research Group and the Public Sector AI Task Force at the Scottish Government Legal Directorate. The views expressed here are personal.