Getty Images v. Stability AI: The Copyright Test

---

The Background: Stable Diffusion and the Image-AI Boom

To understand the Getty v. Stability AI lawsuit, you have to understand Stable Diffusion — what it is, how it was trained, and why its release in August 2022 touched off a firestorm in the visual-arts community.

Stable Diffusion is a generative AI model. Given a text prompt — “a cat sitting on a windowsill at sunset, watercolour style” — it generates an image that matches the prompt. It was developed by a company called Stability AI, in collaboration with researchers at the Ludwig Maximilian University of Munich and the Runway ML research lab. The underlying technology, called a diffusion model, was not new — but Stable Diffusion was the first image-generation model that was both highly capable and openly available. Anyone could download the model, run it on their own hardware, and generate images for free.

This was a significant departure from the existing image-generation models, which were closed and accessible only through paid APIs. OpenAI’s DALL-E, released in January 2021, was available only through OpenAI’s interface. Midjourney, released in July 2022, was available only through a Discord-based interface with a paid subscription. Stable Diffusion, released in August 2022, was open. Within weeks of its release, it had been downloaded millions of times, and the internet was flooded with AI-generated images.

The open release made Stable Diffusion popular. It also made it controversial. The reason is that Stable Diffusion, like all generative AI models, had to be trained on a large dataset of images. The dataset used to train Stable Diffusion was called LAION-5B, and it was built by a German non-profit organisation called LAION. LAION-5B contained 5.85 billion image-text pairs — images collected from across the internet, along with the text that appeared near them. The images in LAION-5B included photographs by professional photographers, illustrations by working artists, screenshots of movies and television shows, and many other kinds of visual content. Much of this content was copyrighted.

The training process worked like this: the model was shown the images in the dataset, along with the associated text, and it learned the statistical relationships between text and images. It learned what a “cat” looks like, what a “watercolour” looks like, what a “sunset” looks like, and how these concepts combine. When you give the model a text prompt, it uses what it has learned to generate an image that matches the prompt.

The controversy was about consent and compensation. The photographers and artists whose work was in LAION-5B had not been asked for permission. They had not been compensated. They had not even been told that their work was being used. And the model that was trained on their work was now being used to generate images that competed with their work — often in their own styles.

This was the context in which Getty Images decided to sue Stability AI. Getty was, in a sense, the perfect plaintiff. It was a large, well-funded company with a massive library of copyrighted images. It had the resources to litigate against a well-funded defendant. And it had a clear, specific grievance: Stability AI had copied millions of Getty photos, with Getty’s watermarks and metadata, and had used them to train a model that could generate images that competed with Getty’s business.

---

The Lawsuits: UK and US

Getty filed two lawsuits against Stability AI, in two different countries, with two different legal theories.

The first lawsuit was filed in the High Court of London (Chancery Division) on January 16, 2023. The UK lawsuit alleged several claims: primary copyright infringement (that Stability AI copied Getty’s photos when it trained Stable Diffusion), secondary copyright infringement (that Stability AI distributed the trained model, which contained reproductions of Getty’s photos), database right infringement (that Stability AI extracted and re-used parts of Getty’s database), trademark infringement (that Stable Diffusion generated images containing Getty’s watermarks), and passing off (that Stability AI misrepresented its products as being associated with Getty). Getty’s primary infringement claim — that Stability AI copied its photos during training — was abandoned when Getty could not prove training occurred in the UK.

Getty’s database-right claim — that Stability AI extracted and re-used parts of Getty’s database — was abandoned by Getty at trial (not rejected by the court), because Getty could not prove that the scraping or training had occurred in the UK. The passing-off claim — that Stability AI misrepresented its products as associated with Getty — was not addressed by the court: Mrs Justice Joanna Smith declined to rule on it, on the basis that it added nothing to her trademark findings and that certain points had not been properly argued. The trademark claim — that Stable Diffusion’s reproduction of the Getty watermark infringed its trademarks — was the claim on which Getty prevailed in the UK, though the court described the finding as “extremely limited” (it applied only to iStock watermarks in early versions of Stable Diffusion; the s.10(3) dilution claim and all Getty-Images-mark s.10(1) claims failed).

The second lawsuit was filed in the United States District Court for the District of Delaware on February 3, 2023 (case 1:23-cv-00135). The US lawsuit alleged copyright infringement, trademark infringement, violations of the Digital Millennium Copyright Act (specifically section 1202, which prohibits the removal of copyright management information), and deceptive trade practices. The US complaint specifically alleged that Stability AI had copied “more than 12 million photographs” from Getty’s collection, along with their captions and metadata.

The two lawsuits were parallel but not identical. The UK and US legal systems approach copyright differently, and Getty’s legal team tailored each lawsuit to the relevant jurisdiction. The UK lawsuit emphasised the secondary infringement claim — the argument that the trained model itself was an infringing copy, because it contained reproductions of Getty’s photos. The US lawsuit emphasised the DMCA claim — the argument that Stability AI had stripped Getty’s copyright management information from the photos when it included them in the training data.

The most striking aspect of the lawsuits was the watermark issue. Stable Diffusion, when generating certain kinds of images, would sometimes produce images that contained what looked like Getty’s watermark — the Getty logo, or fragments of it, superimposed on the generated image. This was a side effect of the training process: the model had seen so many Getty images with watermarks that it had learned to associate the watermark with certain kinds of images, and it sometimes reproduced the watermark when generating similar images. Getty argued that this was trademark infringement — that the generated images, by carrying Getty’s watermark, created the false impression that they were associated with Getty or endorsed by Getty.

---

Stability AI’s Defence

Stability AI’s defence had several components. The details varied between the UK and US cases, but the broad arguments were similar.

First, Stability AI argued that the training of Stable Diffusion did not infringe Getty’s copyright. The argument was that the training process did not reproduce Getty’s photos in any meaningful sense — it extracted statistical patterns from the photos, but it did not store the photos themselves. The trained model, Stability AI argued, was a set of mathematical weights — numbers that represented the model’s learned associations. The model did not contain copies of the training images, and generating an image with the model did not reproduce any specific training image.

Second, Stability AI argued that the training process was fair use (in the US) or fair dealing (in the UK). Fair use — the US doctrine allowing unauthorised use of copyrighted material for transformative purposes — is Stability AI’s central US defence. Fair dealing — the UK analogue of fair use, more limited in scope — was Stability AI’s parallel UK defence. The argument was that training a generative AI model on publicly available images is a transformative use — it does not reproduce the images for their original purpose, but uses them to build a statistical model that can generate new images. This is the same fair use argument that OpenAI has made in the NYT v. OpenAI case, and it is the central legal question in all of the AI copyright cases.

Third, Stability AI argued that the watermark issue was a minor technical problem, not a trademark violation. The argument was that the watermarks appeared in only a small fraction of generated images, that they were the result of the model’s statistical learning rather than any intent to reproduce Getty’s trademark, and that Stability AI was working to eliminate the problem.

Fourth, in the UK case, Stability AI raised a jurisdictional argument. The Copyright, Designs and Patents Act 1988 (CDPA) — the principal UK copyright statute — provides that copyright is infringed by certain acts done “in the UK,” and defines “infringing copy” in section 27. Stability AI argued that the training of Stable Diffusion had not taken place in the UK — it had taken place in other countries, where the servers and the training data were located. Therefore, Stability AI argued, the UK court did not have jurisdiction over the primary infringement claim. This jurisdictional argument would succeed at trial, when Getty could not prove otherwise — but first, the case had to survive a motion to dismiss.

---

The December 2023 UK Ruling: The Case Survives

On December 1, 2023, Mrs Justice Joanna Smith of the UK High Court (Chancery Division) handed down a ruling on Stability AI’s application for strike-out and reverse summary judgment. (The strike-out application had been issued on July 28, 2023 and heard in October 2023.) Justice Smith would later also issue the November 2025 trial judgment. Stability AI had asked the court to dismiss Getty’s claims before trial, arguing that they had no reasonable prospect of success.

The court refused the application. Justice Smith held that Getty’s claims had a reasonable prospect of success and should be allowed to proceed to trial. The ruling was a significant moment in the case, because it meant that the UK case would continue, and that the substantive legal questions — whether training on copyrighted images infringes copyright, whether the trained model is an infringing copy — would be decided at trial.

The ruling was also significant because it addressed Stability AI’s jurisdictional argument. Justice Smith held that the question of whether the training had taken place in the UK was a factual question that could not be resolved on a strike-out application. The question would have to be resolved at trial, on the basis of evidence about where the training had actually occurred.

The case proceeded to trial, which was scheduled for June 2025.

---

The Corporate Side Story: Stability AI’s Turmoil

While the Getty lawsuit was making its way through the courts, Stability AI itself was going through a period of significant corporate turmoil. This turmoil is relevant to the lawsuit because it affected Stability AI’s ability to defend itself, and because it illustrated the broader commercial pressures facing the open-source AI industry.

On March 23, 2024, Emad Mostaque, the founder and CEO of Stability AI, resigned — to pursue “decentralised AI,” as he put it, reportedly saying that “you’re not going to beat centralized AI with more centralized AI.” Mostaque had been a controversial figure — a former hedge fund manager who had positioned himself as a champion of open-source AI. His departure was sudden and unexpected.

Mostaque’s departure was followed by a period of instability. Shan Shan Wong (the COO) and Christian Laforte (the CTO) served as interim co-CEOs from March to June 2024, steering the company through its most acute financial crisis. The company cut 10% of its staff. In the first quarter of 2024, Stability AI reported losses exceeding $30 million on revenues of less than $5 million (per Reuters). The company was, by all accounts, running out of money.

On June 25, 2024, Stability AI appointed a new CEO: Prem Akkaraju, the former CEO of Weta Digital (the visual-effects company known for its work on the Lord of the Rings and Avatar films). Along with Akkaraju’s appointment, the company announced an $80 million funding round (per the Wall Street Journal), which gave the company enough runway to continue operating and to continue defending the Getty lawsuit.

The corporate turmoil at Stability AI was a reminder that the open-source AI business model was, and is, unproven. Open-source AI companies give away their models for free, and they have to find other ways to make money — through enterprise services, through partnerships, through premium offerings. This is difficult, and Stability AI’s struggles were a warning to other open-source AI companies. The financial pressure also affected Stability AI’s ability to litigate. A company that is losing money and cutting staff has fewer resources to defend itself in a multi-year, multi-jurisdiction lawsuit.

---

The June 2025 UK Trial

The UK trial commenced on June 9, 2025. The trial was originally scheduled for 18 days, but it ended up taking 10 actual hearing days, spread across June 9-12, 17-20, 25-27, and 30. The trial was the first in the UK — and one of the first in the world — to address the question of whether training a generative AI model on copyrighted works infringes copyright.

At trial, Getty faced a significant evidentiary problem. To win on the primary infringement claim, Getty had to prove that the training of Stable Diffusion had taken place in the UK. This was the jurisdictional question that Justice Smith had declined to resolve at the strike-out stage. At trial, Getty struggled to produce the evidence. Jones Day, a law firm that published an analysis of the trial, noted that “Getty struggled to come up with the evidence to substantiate its claim and was unable to improve its position in cross-examination of Stability’s employees.”

At the close of the trial, Getty made a significant decision: it abandoned the primary copyright infringement claim. Shepwedd, another law firm that analysed the case, noted that “Getty abandoned its primary copyright infringement claim, accepting that it could not prove the training took place in the UK.” This was a major concession. It meant that the UK court would not decide the most important question in the case — whether training on copyrighted images infringes UK copyright. The court would only decide the secondary infringement claim (whether the trained model was an infringing copy) and the trademark and passing off claims.

The abandonment of the primary infringement claim was a strategic decision, but it had significant consequences. It meant that the UK case would not resolve the central legal question that everyone was waiting for. And it meant that Getty’s primary infringement claim would have to be litigated in another jurisdiction — most likely the United States, where the question of where training occurred is less legally significant.

---

The November 2025 UK Judgment: Model Weights Are Not an “Infringing Copy”

On November 4, 2025, Mrs Justice Joanna Smith handed down a 205-page judgment ([2025] EWHC 2863 (Ch)). The judgment was a surprise, and it was widely seen as a victory for Stability AI — though a partial one.

The central holding of the judgment was on the secondary copyright infringement claim. Getty had argued that the trained Stable Diffusion model was an “infringing copy” of Getty’s photos, because the model contained reproductions of the photos. The court rejected this argument. Justice Smith held that the Stable Diffusion models “do not contain or store reproductions of the relevant works on which they were trained.” The models store mathematical weights — numbers that represent the model’s learned associations — not actual image copies. Therefore, the models are not “infringing copies” under section 27 of the Copyright, Designs and Patents Act 1988, and the secondary infringement claim failed.

This was a significant holding. It was the first time a court in any jurisdiction had directly addressed the question of whether a trained AI model is an infringing copy of its training data. The court’s answer — that it is not — was a major victory for the AI industry, because it meant that distributing a trained model does not, by itself, constitute copyright infringement of the training data.

The judgment was not a complete victory for Stability AI. On the trademark claim, the court found in favour of Getty, holding that the generation of synthetic images with Getty watermarks infringed Getty’s UK trademarks. This was a limited finding — it applied only to images that actually contained the watermark, not to all images generated by Stable Diffusion — but it was a real finding of liability.

The passing-off claim was not addressed on the merits — the court declined to rule on it, on the basis that it added nothing to the trademark findings and that certain points had not been properly argued. The database-right claim had been abandoned by Getty at trial (because it could not prove UK-located scraping), so the court did not rule on it either.

The most important thing about the judgment was what it did not decide. Because Getty had abandoned the primary infringement claim, the court did not decide whether training on copyrighted images infringes UK copyright. The court explicitly noted this, leaving the larger question for another day. Katten, a law firm that analysed the judgment, noted that it “leaves critical questions unanswered relating to the relationship between IP rights and generative AI, particularly whether the use of copyright works for training infringes.”

Paul Weiss, another law firm, made a pointed observation about the implications: “Critically, the case implies that primary infringement of UK copyright works could be evaded by training models in other jurisdictions.” This is a significant implication. If a company can avoid UK copyright liability by training its models outside the UK, then the UK’s copyright law provides less protection for UK creators than might have been assumed. This is a problem that the UK Parliament may need to address.

Getty was granted leave to appeal on December 16, 2025 (per the ICLR record, [2025] EWHC 3343 (Ch)). The appeal will focus on the secondary-infringement holding (the model-weights ruling) and on the trademark finding. The appeal will not, however, be able to revive the primary infringement claim that Getty abandoned at trial.

---

The US Case: Still Pending

While the UK case was proceeding to judgment, the US case was proceeding on a different track. On August 14, 2025, Getty voluntarily dismissed its Delaware action (1:23-cv-00135) and refiled the same day in the US District Court for the Northern District of California (3:25-cv-06891). The refiling was a strategic move: the Northern District of California is a more favourable venue for copyright plaintiffs, and it is also where many of the parallel AI copyright cases are being heard.

On April 23, 2026 (some sources cite April 24, reflecting order-entry vs. filing discrepancies), Judge Trina L. Thompson of the Northern District of California denied most of Stability AI’s motion to dismiss. (Judge Thompson was assigned the case after Getty’s August 14, 2025 refiling from Delaware.) The surviving claims included direct copyright infringement, trademark infringement, false designation of origin, and California unfair competition. The only claim that was dismissed — and only without prejudice, meaning it could be repleaded — was the DMCA section 1202(a) claim for providing false copyright management information.

The US case is now in the discovery phase — the pre-trial evidence exchange — as of mid-2026. No trial date has been set. The US case will address the central question that the UK case did not: whether training on copyrighted images, without the permission of the copyright holder, is fair use under US law. This is the same question that is at issue in the NYT v. OpenAI case (see B81), and it is the most important unresolved legal question in the AI industry.

The US case will also address the question of where training occurred, which is less legally significant in the US than in the UK. Under US law, copyright infringement occurs when the infringing act takes place in the US, regardless of where the training servers are located. This means that Stability AI will not be able to make the same jurisdictional argument in the US that it made in the UK.

---

What the Case Means

The Getty v. Stability AI case has already had significant implications, and it will continue to have implications as the US case proceeds.

The most important implication of the UK judgment is the holding that model weights are not an “infringing copy” of the training data.

This is a significant victory for the AI industry, because it means that distributing a trained model does not, by itself, constitute copyright infringement. AI companies can distribute their models without fear that the distribution itself infringes the copyright of the creators whose work was used in training. This does not resolve the question of whether the training itself infringes — that question remains open — but it does remove one theory of liability that could have been very damaging to the AI industry.

The second implication is that the primary infringement question — whether training on copyrighted works infringes copyright — remains unresolved in the UK. Getty’s decision to abandon the primary infringement claim, because it could not prove that training occurred in the UK, means that the UK courts have not addressed the most important legal question. This question will have to be decided in another case, or in another jurisdiction. The US case will likely be the first to address it directly.

The third implication is that the trademark claim — the watermark issue — is a real liability for image-generation AI companies. The UK court’s finding that generating images with Getty watermarks infringes Getty’s trademarks means that image-generation AI companies need to take steps to prevent their models from reproducing watermarks. This is a technical problem that can probably be solved, but it is a real legal risk that needs to be addressed.

The fourth implication is that the corporate turmoil at Stability AI illustrates the broader commercial pressures facing the open-source AI industry. Open-source AI companies are giving away their models for free, and they are struggling to find sustainable business models. The financial pressure makes it harder for them to defend themselves in litigation, and it makes the outcome of the litigation more consequential — a company that is already struggling may not survive a significant legal defeat.

The fifth implication is that the case has revealed a gap in the UK’s copyright law. The fact that a company can avoid UK copyright liability by training its models outside the UK is a problem that the UK Parliament may need to address. Other countries may face similar gaps, and the case is a reminder that copyright law — which was designed for a world of physical copies and discrete works — is struggling to adapt to a world of AI training and statistical models.

---

Where the Case Stands

As of mid-2026, the UK case has been decided, and Getty has been granted leave to appeal. The appeal will likely be heard in 2026. The US case is in the discovery phase, with no trial date set. The US case will address the central question of whether training on copyrighted images is fair use under US law.

The case is, in a sense, the image-AI counterpart to the NYT v. OpenAI case. Both cases turn on the same fundamental question — whether training AI models on copyrighted content, without the permission of the copyright holder, is legal. But the cases are proceeding on different timelines and in different jurisdictions, and they may produce different answers. The UK case has already produced one significant holding — that model weights are not an infringing copy — that the US case will need to consider.

The outcome of the US case will have significant consequences for the image-generation AI industry, and for the broader AI industry. If the US court holds that training on copyrighted images is fair use, it will validate the data-acquisition model that the image-generation companies have been using, and it will allow them to continue building models using the enormous amounts of image data available on the internet. If the US court holds that training on copyrighted images is not fair use, it will require image-generation companies to license the images they use for training, which will be expensive and will give large image libraries like Getty significant leverage.

The case is also a reminder that the legal framework for AI is still being worked out. The questions that the case raises — about training, about model weights, about trademarks, about jurisdiction — are new questions, and the legal system is still figuring out how to answer them. The answers will shape the development of the AI industry for years to come.

---

---

The story of getty images v. stability ai is not an abstraction. It is happening in the products you use, the systems that govern your life, and the institutions you rely on. Understanding it is the first step. Demanding accountability is the second.