Getty v. Stability (UK)

What Getty alleged

Getty sued Stability AI in London in January 2023, alleging Stable Diffusion was trained on millions of Getty photographs scraped without a licence, and that its outputs reproduced both the substance of those images and, in some cases, distorted versions of Getty’s and iStock’s watermarks.

The claim ran on parallel tracks with Getty’s separate US case against Stability, filed around the same time — a strategy of pursuing the same underlying conduct in two jurisdictions with different copyright frameworks, betting that at least one forum would produce a ruling on the merits of AI training.

By the time the case went to trial in June 2025, it had become the most closely watched AI copyright dispute in the world, largely because England’s Chancery Division was on track to be the first major court anywhere to hear a full trial — with live witnesses and cross-examination — on whether training a generative model infringes copyright.

The jurisdictional collapse: why the training claims were never decided

When trial opened on June 9, 2025, evidence about where Stable Diffusion’s training actually took place became central to the case. By June 25, Getty withdrew its primary copyright claims over training and outputs, after the evidence indicated the training occurred outside the UK — putting those acts beyond the territorial reach of UK copyright law.

That withdrawal meant the question most closely watched by the industry — whether training an AI model is itself an infringing act under UK copyright law — was never decided on the merits. Getty’s own evidence undercut UK jurisdiction over the conduct before the court could rule on it, leaving only the narrower secondary-infringement and trademark theories for the judge to resolve.

The ruling and its limits: secondary infringement rejected

In her November 4, 2025 judgment, Mrs Justice Joanna Smith dismissed Getty’s remaining secondary-infringement theory. She held that a model which never stored or reproduced the underlying copyright works does not qualify as an “infringing copy” under section 27 of the Copyright, Designs and Patents Act — meaning the model weights themselves, as an intellectual product, don’t count as a copy of the training images for this purpose.

Getty’s win was correspondingly narrow: trademark findings limited to watermark artifacts that appeared on outputs of early Stable Diffusion versions, described in the judgment as largely historic. There was no broader finding on the copyright merits of AI training.

“A model which never stored or reproduced the underlying copyright works is not an “infringing copy” within the meaning of section 27 of the Copyright, Designs and Patents Act 1988.”— Judgment, [2025] EWHC 2863 (Ch) (Nov. 4, 2025)

Where it stands: appeal granted

At a consequentials hearing on December 16, 2025, the court granted Getty permission to appeal the secondary-infringement dismissal in full, with the appeal due to be filed by February 3, 2026. No Court of Appeal decision has been reported as of July 22, 2026.

The appeal is confined to the section 27 “infringing copy” question the court actually decided — not to the training-jurisdiction issue, which Getty abandoned when it withdrew those claims mid-trial. That narrows what the Court of Appeal can ultimately resolve, even if Getty prevails.

Why it matters for training-data buyers

The first full trial of an AI training dispute anywhere in the world ended without a ruling on the core question, because the training happened outside the UK. Where a model is trained now determines which country’s law can reach it — a jurisdictional reality that shapes exposure as much as the substance of copyright law does.

Buyers of models trained abroad cannot assume UK law on this point is settled: if Getty’s appeal succeeds, importing model weights trained on unlicensed data could itself infringe UK copyright, even without in-UK training. Separately, the trademark findings show that watermark remnants surviving into outputs create liability of their own, independent of the copyright question — making provenance and watermark hygiene a supplier-side obligation regardless of how the appeal comes out.

For anyone assessing chain of title on a model trained overseas, this case is a reminder that “where” is now a legal fact worth documenting, not an operational detail — a model’s training venue may determine which regime of copyright law, if any, ever gets to rule on how it was built.