90 incidents on record · 2026 Headlights Incident reports by Ellie Harris · Melbourne
10 new this week Library last updated 13 July 2026
← The incident library
HD-INC-085
Technology · United Kingdom · 2025 · Untraceable training-data provenance

Getty Images sued Stability AI over the images behind Stable Diffusion, then abandoned its main copyright claim because it could not establish where the training happened

By Ellie Harris · Filed Claims filed in the United States and United Kingdom in 2023; UK trial June 2025

Alleged: Stability AI Ltd; Getty Images (US), Inc. (claimant) developed or deployed the AI system implicated in this incident. Details are drawn from public reports; parties are presumed innocent of any wrongdoing not established by an official finding.

Getty Images sued Stability AI over the images behind Stable Diffusion, then abandoned its main copyright claim because it could not establish where the training happened

What happened

It was reported that in 2023 Getty Images brought proceedings against Stability AI, the developer of the Stable Diffusion image generator, in both the United States and the United Kingdom. Getty alleged that Stability had copied more than twelve million of its images, without a licence, to train the model, and that Stable Diffusion had at times reproduced a distorted version of the Getty watermark in the pictures it generated. The UK case came to trial in June 2025 and became one of the first in the country to test how copyright and trade mark law apply to the building of a generative AI model.

The trial did not resolve whether training an AI model on copyrighted images infringes copyright, because the claim that would have tested it was withdrawn. Shortly before closing submissions Getty dropped its primary copyright infringement and database right claims, accepting there was no evidence that the training and development of Stable Diffusion had taken place in the United Kingdom, which is where the law required those acts to have happened for the claim to succeed. On 4 November 2025 Mrs Justice Joanna Smith handed down judgment. She dismissed Getty’s remaining secondary copyright claim, holding that the model did not store or reproduce the copyrighted works and so was not an “infringing copy” under UK law. She found trade mark infringement only in an extremely limited way, against historic versions of Stable Diffusion that had reproduced the Getty watermark, and dismissed the wider trade mark claim; the passing off claim brought Getty no wider finding of liability. Getty was subsequently granted permission to appeal the dismissal of its secondary copyright claim, in December 2025, while Stability was refused permission to appeal the trade mark findings. The separate United States proceedings remained on foot. Stability, in other words, largely prevailed, though much of the case still turned on questions of fact no record had been kept to answer: what the model had been trained on, and where.

What an auditable version would have shown

Two of the questions that shaped the case are questions a record answers directly. The first is provenance: which images were in the training data, drawn from what sources, and under what licence. The second is location: where the training computation actually ran, the fact whose absence forced Getty to drop its main claim. An auditable version treats both as part of the model’s build record: a signed manifest of the training corpus and its sources and licences, and a log of where and when the training was performed. With that manifest, the licensing question is not litigated through inference from watermarks in outputs; it is read from the record of what went in. And the jurisdictional question that ended Getty’s primary claim, where the training happened, is a logged fact rather than an evidentiary gap. The same records would show what recognisable third-party content, a watermark among it, the model was capable of emitting, so that its appearance in an output is a known and controllable property rather than a surprise produced in evidence.

Where the gap was

The building of the model left no record that could establish what it had been trained on, or where. That did not decide the case, but it shaped it: the factual questions a record answers, which images were used and in which country the training ran, had to be fought through inference instead, and it was the second of these, jurisdiction, that forced Getty to abandon its main claim. A ConductRecord applied to model development supplies that record: a signed, reviewable log of the training corpus, its sources and licences, and the location and time of the training runs, so that the provenance and jurisdiction of a model are facts the builder can produce rather than facts a court must reconstruct. It does not answer the legal question the trial left open, whether training on copyrighted work infringes; it removes the evidentiary uncertainty that stopped that question from even being reached. An EgressGate addresses the watermark: it classifies generated outputs for recognisable third-party marks and content before they leave the system, so that reproducing a protected mark is caught and controlled at the point of output rather than surfacing as evidence of what the model absorbed.

What governance should have looked like

Much of this case turned on evidence rather than law. The legal questions were real, and a record would not have decided them, but a large part of the dispute, what the model was trained on and where that training happened, was an evidentiary problem, and evidence is what a build record provides. When the provenance of a training set and the location of the work that used it are not written down, every later question about them, licence, jurisdiction, consent, becomes a matter of reconstruction and argument, expensive for the rights-holder and no more comfortable for the developer, who must defend against claims it also cannot cleanly disprove. A model’s training data is not an incidental detail of its construction; it is the thing most likely to be asked about, by a court, a regulator or a licensor, and the cheapest time to be able to answer is while the model is being built. Best practice would be for a developer that trains a model on large quantities of others’ material to be able to produce a signed record of what it used, from where, under what terms, and where the training ran, so that questions this litigation could only approach by inference could instead be answered from the record.

Failure Pattern: a dispute over what an AI model was trained on could not be resolved from any record, so questions of licence, and even of where the training took place, turned on inference rather than evidence.

Governance Principle: what a model was trained on, sourced from where and under what licence, and where that training took place, should be a signed record the builder can produce, not a fact that must be inferred years later in court.

The reference implementation of ConductRecord and EgressGate is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.

Sources

The mailing list

Fresh incident reports every week. One email to match.

We add new incidents to the library regularly, and send a single short email each week with what's new. The library stays free and open; this is just how you keep up with it.

No tracking. Unsubscribe in one click.

The record

An auditable system would have produced a signed, tamper-evident record the moment this happened: what the system did, the version that did it, the basis it acted on, and the action taken, and Stability AI Ltd; Getty Images (US), Inc. (claimant) could have produced it on demand.

This is the record the system as deployed did not produce in a signed, auditable form.

What this teaches
Capture what happened when it happens
What the system did, the version that did it, the basis it acted on, and the action taken, recorded at the moment, not reconstructed after.
Sign it, so no one has to trust the record-keeper
A tamper-evident entry. Edit it later and the signature breaks. The record does not ask for the benefit of the doubt.
Make it verifiable by anyone
A court, a regulator, a customer's lawyer can check the record themselves, without taking the company, or us, at our word.

Headlights summarises publicly reported AI incidents. All summaries are independently written, attributed to their original sources, and intended for research and educational purposes. Allegations are identified as such until established through official findings.

Last reviewed June 2026. This report is based on the sources listed above and reflects information available at the time of review; later developments may not be captured. Where a person is described as charged with or alleged to have done something, that allegation is unproven unless a conviction or a court or regulatory finding is stated. Headlights publishes journalism and commentary, not legal advice.

Want to write back?

Direct to my inbox.

ellie@useheadlights.com →