Getty v Stability AI: The Court of Appeal Question That Could Redraw UK Copyright for Every AI Model

The Judgment Under the Microscope

Getty Images (US) Inc & Ors v Stability AI Limited [2025] EWHC 2863 (Ch), High Court of England and Wales, 4 November 2025. That is the case most of the AI industry watched. The headline read as a Stability AI victory. The reality, as with most headline-grabbing IP judgments, was considerably messier.

Mrs Justice Joanna Smith dismissed Getty's secondary copyright infringement claim under sections 22 and 23 of the Copyright, Designs and Patents Act 1988. The core finding: the Stable Diffusion model, trained overseas on millions of Getty images without authorisation, did not constitute an 'infringing copy' of 'an article' when imported and distributed in the UK. The model weights, the court held, did not reproduce the protected works - they merely encoded statistical relationships derived from them. Getty had already abandoned its primary copyright and database rights claims before trial, because it could not establish that training or development had taken place in UK jurisdiction. That evidential gap, as practitioners at VWV observed, demonstrates the structural enforcement problem rights holders face: transparency is a precondition for any infringement claim, and without knowing which works trained a model, the claim cannot be pleaded.

The only partial win for Getty was on trade mark infringement, limited to historic Stable Diffusion versions 1.x and 2.x that generated images bearing a distorted Getty watermark.

The Academic Challenge

Now the judgment faces something almost as significant as the Court of Appeal: academic demolition. The Journal of Intellectual Property Law & Practice, published by Oxford University Press, carried a peer-reviewed article this month contending that the High Court was simply wrong on secondary copyright infringement. The article argues the court incorrectly required an 'infringing article' to consist of or contain a copy of the protected work, contrary to both the statutory language of the CDPA and established case law. The piece closes by examining what this means for the appeal and for the broader regulation of generative AI in the UK.

This matters because academic critique at this level, in this journal, gets read by Court of Appeal judges. It shapes the written submissions of senior counsel. It is not decoration; it is input.

What the Appeal Will Actually Decide

Mrs Justice Smith granted Getty permission to appeal in full at the consequentials hearing on 16–17 December 2025. She observed that the question of whether an 'infringing article' can be an intangible item had not been considered by any UK court previously, and that Getty's proposed appeal had "a real prospect of success" on a "pure question of law" involving "statutory construction on which the minds of reasonable lawyers may differ."

So: the Court of Appeal must decide whether model weights, delivered remotely or imported as intangible digital assets, can constitute 'infringing articles' under ss. 22–23 CDPA. If it decides yes, the consequences cascade. An AI model trained overseas on unlicensed copyright works would be legally toxic the moment it crossed into UK jurisdiction, whether by download, API call, or cloud deployment. Every AI provider offering services into the UK market would need to audit their training data provenance or risk secondary infringement liability on the basis of importation alone.

Counsel for the respective parties are likely to fight hard on the question of what 'article' means in a statute drafted in 1988, before the commercial internet existed. The statute wasn't drafted for this technology, and it shows. The legislature had physical articles in mind - bootleg CDs, counterfeit prints. Whether Parliament's silence on digital intangibles should be read as exclusion or as interpretive latitude is precisely the kind of question where the Court of Appeal earns its authority.

Noteworthy: Stability was refused permission to appeal the trade mark findings. Those narrow holdings - watermark infringement on legacy model versions - stand.

The Policy Vacuum That Created This Litigation

The Getty appeal does not exist in isolation. It is a direct consequence of legislative paralysis. On 18 March 2026, the UK Government published its Report on Copyright and Artificial Intelligence, laid before Parliament under sections 135–137 of the Data (Use and Access) Act 2025. It resolved nothing. The government confirmed it will not introduce a broad text and data mining exception for commercial AI training - abandoning the opt-out mechanism it had previously signalled as its preferred approach. It equally declined to introduce mandatory licensing. The existing framework under section 29A CDPA, which permits copying for non-commercial research only, remains the operative law.

The House of Lords Communications and Digital Committee, which published its own report on 6 March 2026, had urged a 'licensing-first' approach, rejecting any broad TDM exception. The government's May 2026 response to that committee confirmed the retreat from the opt-out mechanism. Creative industries had mobilised 11,500 consultation responses, analysed manually by around 80 government officials - no AI involved, the UCL Copyright Blog noted with some irony. Rights holders won the political argument. But winning the political argument and winning the legal argument are different things. As VWV's analysis observed bluntly: policy has shifted from Parliament into the courts and the commercial marketplace, at least for now.

So the Court of Appeal is being asked to fill a void that Government explicitly declined to fill.

The Brand Signal in the Filing Data

Here the broader market context sharpens the stakes. AIBD analysis of UK IPO trademark data shows 6,503 Class 35 filings in 2026-Q3, a fall of 39.7% against the comparable prior period. Class 35 covers advertising, business management, and retail services - the commercial heartland of AI product and platform brands. A contraction of that magnitude in a single quarter, across the class most commonly used by AI-adjacent businesses to protect brand identity, suggests that at least some market participants are pausing brand investment while the legal framework for underlying IP remains unresolved. You do not file a trademark to protect a business whose copyright foundation may shift materially on appeal. The filing numbers do not prove causation. But the timing is not coincidental.

TrademarkDashboard data confirms that Class 35 and Class 9 (software and electronics) are among the most heavily filed classes in the UK register, which holds approximately 3 million records and receives 70,000–90,000 new applications annually. A near-40% drop in one quarter is a material deviation.

The Historical Parallel

The closest structural precedent is Amstrad Consumer Electronics plc v The British Phonographic Industry Ltd [1986] FSJ 159, in which the Court of Appeal considered whether a manufacturer of dual-cassette decks with high-speed dubbing was liable for enabling consumers to infringe copyright. The House of Lords ultimately found no authorisation liability. But the case dragged on for years and rewrote commercial behaviour in the consumer electronics sector long before a final ruling. Companies altered product specs, licensing practices, and marketing - not because the law was clear, but because it was not. The same dynamic is operating now, at ten times the commercial velocity.

What Comes Next

Getty filed its notice of appeal by the 3 February 2026 deadline. Court of Appeal scheduling in complex IP matters currently runs to approximately twelve months from permission, which puts the likely hearing window in early-to-mid 2027. Expect both sides to seek expedition given the commercial urgency. The JIPLP article published this month may feature in written submissions if it is formally cited by either party's counsel.

In the meantime, AI developers distributing models into the UK market face an uncomfortable interregnum. The High Court said their model weights are not infringing articles. The Court of Appeal may say otherwise. Prudent legal advice - and several major firms are giving precisely this - is to document training data provenance now, pursue licences for any commercially significant datasets, and treat the High Court ruling as a speed bump, not a wall.

The UK's position as a jurisdiction of choice for AI litigation is being built, case by case, in exactly this way. That is not necessarily a bad thing. But it is an expensive thing, and the bill is being paid by creative industries and AI developers alike while Parliament watches from the public gallery.