Breaking
Loading headlines…
AI Business Dispatch.

Getty v. Stability AI Reaches the Court of Appeal: Can a Model Be a Copy of Works It Does Not Contain?

The UK's most important AI copyright case is now before the Court of Appeal, with a central question that existing copyright law was never designed to answer - and a ruling that will shape how every AI developer in Europe approaches training data.

E
Eleanor Vance-Hartley · 15 November 2026 · 6 min read
Share:·X·LinkedIn
Getty v. Stability AI Reaches the Court of Appeal: Can a Model Be a Copy of Works It Does Not Contain?
Eleanor Vance-Hartley

Court of Appeal, England and Wales - November 11-12, 2026

Getty Images (US), Inc. v. Stability AI Ltd., Court of Appeal of England and Wales, CA-2026-000007. The hearing ran over two days, 11 and 12 November. No judgment date has been set.

What was actually decided at first instance matters here, because the headlines still get it wrong. Mrs Justice Joanna Smith's November 2025 judgment did not hold that AI training is lawful in the UK. It held something narrower, and in some ways more damaging to rights holders: that the Stable Diffusion model weights do not constitute an "infringing copy" under the secondary infringement provisions of the Copyright, Designs and Patents Act 1988. The primary copyright claim, whether training itself constitutes infringement, was never adjudicated. Getty, represented by Fieldfisher, had already dropped the core copyright part of its claim during the trial in June, chiefly because it could not establish a strong enough jurisdictional link between the UK and the training of Stability AI's model, which took place abroad. The statute was not drafted for this technology, and it shows.

Advertisement

TrademarkDashboard: Stop counterfeits before they hijack your brand.

What Getty Is Arguing on Appeal

Getty is appealing the High Court's November 2025 ruling that largely favoured Stability AI, arguing that an AI model trained overseas on unlicensed works can still be an "infringing copy" when imported into the UK. The legal peg is secondary infringement under the CDPA 1988: specifically, whether a model constitutes an "article" that can be an infringing copy, and whether making Stable Diffusion available in the UK constitutes importation or distribution of that article.

Stability AI is also asking the Court of Appeal to overturn two first-instance findings that went against it: that an intangible model can be an "article", and that Stability AI itself imported or distributed the model in the UK. Both sides are cross-appealing aspects of the first-instance decision. The court must somehow apply mid-twentieth century statutory language, drafted when "articles" meant physical things, to billion-parameter model weights existing as floating-point numbers on a server.

Getty has asked the appeals court to overturn the trial loss of its secondary copyright infringement case against Stability AI, with a trade group warning that UK copyright law should not enable "offshore laundering" of copyright material. That framing is deliberate. It maps precisely onto the argument gaining traction in the United States, where the Third Circuit's September 2026 decision in Ross Intelligence found the market for licensing content as AI training data "rapidly developing," a finding that weakens the fair-use defence for companies that bypass licensing deals. The wind is at Getty's back, jurisprudentially speaking. Whether the Court of Appeal agrees is another matter.

Why the Judgment Will Disappoint

Fieldfisher and Bird & Bird are the rival IP teams back in court for this closely watched appeal, but observers hoping for answers to some of the technology's biggest legal questions may be left disappointed. That assessment from Managing IP, published on 11 November, is correct. The appeal is bounded by its procedural history. Because the primary infringement question, whether UK-based training of an AI on copyright works constitutes infringement at all, was abandoned before judgment, a core question about UK copyright law remains unanswered: does training in the UK of an AI model on copyright-protected works constitute infringement under UK law?

The Court of Appeal cannot answer a question that was never put to it. Whatever it decides, that foundational gap in UK AI copyright law will persist until either Parliament acts or a future claimant manages to run the primary infringement argument to judgment.

If the appeal fails, the training question goes back to Parliament, where the pressure for reform is already building. The government's March 2026 statutory report under s.137 of the Data (Use and Access) Act has not resolved the tension between the creative sector and AI developers. A Court of Appeal loss for Getty would increase that pressure considerably.

The Intermediary Problem: Brave Software and the Supply Chain

This UK appellate hearing did not arrive in a vacuum. Two weeks earlier, on 2 November, the New York Times, New York Daily News, Chicago Tribune, and Orlando Sentinel sued Brave Software in federal court, accusing the web browser and AI search company of illegally copying their reporting and selling access to copyrighted content to artificial intelligence companies, including Anthropic and Perplexity.

The Brave litigation is structurally different from the direct-training cases, and that difference matters. It challenges the assumption that retrieving content through an established search index or third-party API somehow eliminates the underlying copyright problem. Both cases target the Brave Index, a more than 30-billion-page web index that feeds Brave Search and is sold via API to Anthropic, Perplexity, Mistral, xAI, and Cohere. The complaints allege direct, contributory and vicarious copyright infringement, DMCA circumvention, and unjust enrichment.

This is the intermediary-liability theory applied to AI supply chains: Brave as the unlicensed data distributor, the AI labs as the knowing receivers. Both complaints call Brave "the supplier, dealer, and distributor of stolen copyrighted content" and its AI customers "knowing receivers." Whether courts treat the Brave Search API as more analogous to a photocopier, a news wire, or a licensed syndication desk will determine whether the claim survives summary judgment.

One detail from the complaints deserves to stand alone: in one test, Brave's API returned 1,291 of a 1,295-word article. That is not retrieval. That is reproduction.

The UK Trademark Filing Signal

Against this backdrop of contested copyright boundaries, trademark registration data reveals an uncomfortable divergence. AIBD analysis of UK Intellectual Property Office data shows that Class 35 filings, covering advertising, business management and retail services, the category most relevant to AI-platform commercial activity, came in at just 5,310 in Q4 2026, a fall of 53.1% against the prior period. Class 35 is where AI-enabled service businesses should be registering their brand protection; the collapse in filings suggests that legal uncertainty around AI outputs is suppressing, not stimulating, brand formalisation activity.

The connection is not coincidental. When the legal framework for the underlying product is contested at appellate level, brand registration tends to stall. Companies do not invest in protecting a commercial identity built on a foundation that might shift beneath them. The trademark filing data is a downstream symptom of upstream copyright dysfunction.

The Historical Parallel

The closest analogy in copyright history is the cable television disputes of the 1970s, where US courts grappled with whether cable operators retransmitting broadcast signals were liable for infringement or merely passive conduits. The Supreme Court's decisions in Fortnightly Corp. v. United Artists (1968) and Teleprompter Corp. v. CBS (1974) held that cable retransmission was not a "performance" under the 1909 Act, effectively forcing Congress to legislate. The 1976 Copyright Act created a compulsory licensing scheme that resolved the impasse. AI copyright is on the same trajectory. Courts are applying statutes that were not drafted for this technology, delivering decisions that disappoint everyone, and Parliament, like Congress fifty years ago, will eventually have to act.

What Happens Next

The Court of Appeal's judgment has no date given. Realistically, given the complexity of the cross-appeals and the absence of binding UK authority on these precise points, a ruling before Q2 2027 would be optimistic. Meanwhile, the outcome will influence how AI developers approach training data licensing across Europe and could inform legislative efforts to clarify copyright boundaries for generative AI systems.

In the United States, the Brave Software litigation will work through the Southern District of New York on a track that could produce a summary judgment ruling by late 2027. The Ross precedent from the Third Circuit makes Brave's fair-use defence harder to run; the company has previously defended its search practices as fair use, and had earlier sued News Corp seeking a ruling that its search practices were lawful. That pre-emptive suit now reads as a miscalculation.

Three tracks are live simultaneously: a UK Court of Appeal that will clarify secondary infringement but cannot touch the primary question; a US district court testing whether intermediary copyright liability extends to API-based content distribution; and a UK Parliament still searching for the legislative formula that balances developer access against creator compensation. None of them moves fast. The law, as it stands today, is insufficient for the technology it is being asked to govern.

ip-lawai-copyrightpatentstrademarksgetty-imagesstability-aibrave-softwarecdpa-1988secondary-infringementdmcatraining-datauk-court-of-appealdata-scraping