Getty Images v. Stability AI: What the UK and US Cases Mean for AI Image Developers and Users

If your business uses AI image generation tools — whether to produce marketing materials, social media graphics, product visuals, or website illustrations — you need to understand what is happening in two courtrooms on opposite sides of the Atlantic. The litigation between Getty Images and Stability AI, the developer of the Stable Diffusion image generation model, has produced the first judicial ruling in the United Kingdom on whether training an AI model on copyrighted images is lawful, and it has left a parallel American case actively pending in California. Together, these proceedings are drawing the first rough legal boundary lines around AI-generated imagery, and those lines will affect every company that uses these tools commercially.

This post walks through what was alleged, what was decided, how UK and US copyright law diverge on the underlying question of AI training data, and what practical steps your business should be taking right now.

Background: What Getty Images Alleged

Getty Images is one of the world’s largest stock photography agencies, licensing photographs and illustrations to publishers, advertisers, and businesses across the globe. Its business model depends on controlling who can use the images it owns or exclusively licenses, and charging accordingly.

Stability AI built and commercialized Stable Diffusion, a text-to-image generative AI model. Type a text prompt, and the model produces a photorealistic or artistic image on demand. The company launched Stable Diffusion as an open-source model in 2022 and built a commercial product, DreamStudio, on top of it. The model was trained on the LAION-5B dataset, an enormous collection of image-text pairs assembled by LAION, a German non-profit that received funding from Stability AI. LAION built its dataset by crawling the open web — specifically by processing Common Crawl data, collecting images paired with alt text from billions of HTML pages. One analysis found that approximately 12 million of the images in the training data came from Getty’s websites.

Getty alleged that Stability AI scraped those images without a license, without payment, and without any meaningful authorization, and then used them to train a commercial product that directly competes with Getty’s own licensing business. The core theory was straightforward: Stability AI took Getty’s product, used it to build a competing product, and sold that competing product to Getty’s own customers.

Getty filed separate proceedings in the United Kingdom, where Stability AI is incorporated, and in the United States.

The Watermark Problem: Evidence of Copying

One of the most striking pieces of evidence in both cases was not a legal argument — it was a screenshot. Users discovered that Stable Diffusion, when prompted with terms like “stock photo,” “editorial photograph,” or certain descriptive phrases, occasionally generated images bearing distorted versions of the Getty Images or iStock watermarks. These are the semi-transparent logos that Getty and iStock overlay on preview images to prevent unauthorized use.

Independent investigators confirmed the phenomenon. When prompts included the phrase “stock photo,” the model would frequently produce images with watermark-like markings that closely resembled Getty’s branding. The effect was reproducible without any contrived or unusual prompting.

Technically, this happened because the LAION dataset contained millions of watermarked Getty preview images. Although Stability AI filtered for images with a predicted watermark probability below a certain threshold when assembling the LAION-Aesthetics training subset, a substantial number of watermarked Getty images remained in the training data. Through a process that experts described as “memorization” or “overfitting,” the model learned to associate certain visual and textual characteristics — including the watermark itself — with a particular category of image output.

Expert witnesses in the UK case agreed that the watermarks appeared in outputs because the model had been trained on images containing them. They also agreed that in earlier versions of the model (versions 1.x and 2.x), watermarked outputs could be generated with high frequency under realistic, non-contrived prompting conditions. Later versions incorporated filtering measures that materially reduced but did not completely eliminate the phenomenon.

This watermark evidence served a dual purpose in the litigation. On the copyright side, it supported the argument that the model had absorbed and reproduced recognizable features of Getty’s images. On the trademark side, it was far more potent: generating images that display a competitor’s brand mark is a direct and visible form of trademark confusion.

The UK High Court Ruling: [2025] EWHC 2863 (Ch)

On November 4, 2025, the High Court of England and Wales issued the first judicial decision in any jurisdiction addressing whether training a generative AI model on copyrighted images constitutes copyright infringement. The judgment runs to hundreds of paragraphs and is worth understanding in some detail, because it illustrates how the legal question of “was there copying?” does not map neatly onto the technical process of neural network training.

The Primary Copyright Claim: Abandoned

The most significant procedural development in the UK case was one that received relatively little attention: Getty withdrew its primary copyright infringement claim before trial. Primary infringement under UK law requires showing that the defendant itself made a copy of the protected works on UK soil. Getty could not demonstrate that the acts of scraping and processing its images — the training pipeline — occurred within the United Kingdom. Stability AI’s training operations had taken place in the United States, not the UK.

This meant the court never actually decided the central question that this litigation was expected to answer: whether training a generative AI model on copyrighted images, as a matter of substantive UK copyright law, constitutes infringement. That question remains legally open in the UK.

Secondary Copyright Infringement: Rejected

Getty’s secondary copyright infringement claim did proceed. Under the Copyright, Designs and Patents Act 1988 (CDPA), a person infringes copyright by possessing or dealing with an “infringing copy” of a work in the course of business. An “infringing copy” is an article the making of which constituted an infringement. Getty argued that Stable Diffusion’s trained model weights constituted an infringing copy of its images — in other words, that the model itself contained Getty’s works in stored form.

The court rejected this argument. Mrs. Justice Joanna Smith held that the Stable Diffusion model weights do not contain or store reproductions of the images on which the model was trained. The weights are numerical parameters — billions of floating-point values encoding statistical relationships learned during training. The court found that these parameters do not embody any recognizable reproduction of the training images, and thus do not meet the legal definition of an “infringing copy” under the CDPA, which requires that the article itself constitute or contain a reproduction of the protected work.

This was a significant win for the AI development community. The court effectively held that the act of distilling billions of images into a trained neural network does not create a stored copy of those images in any legally cognizable sense.

Trademark Infringement: Getty Partially Prevailed

Where Getty did win was on trademark grounds. The court found that early versions of Stable Diffusion (versions 1.x and 2.x) infringed Getty’s and iStock’s registered trademarks when they generated outputs displaying the Getty watermark for UK users. The court’s reasoning was that outputs bearing the Getty mark created a likelihood of consumer confusion about the source or authorization of the image — a core element of trademark infringement.

However, the court declined to award enhanced or additional damages. The infringement was characterized as historic and limited: it was confined to earlier model versions, it did not appear to reflect deliberate policy on Stability AI’s part, and later versions had taken steps to reduce the problem. The trademark win was real but relatively modest in practical terms.

Stability AI sought permission to appeal the trademark findings and was refused at the trial court level. Getty sought permission to appeal the copyright findings and was granted it. As of this writing, Getty has until early 2026 to file its formal appeal on the secondary copyright question.

The fundamental issue on appeal is whether an “article” can constitute an “infringing copy” even if it does not physically contain or store copies of the protected works. The Court of Appeal’s answer will either affirm the lower court’s model-weights analysis or potentially reopen the question of how copyright law applies to trained AI systems.

The US Proceedings: From Delaware to California

Getty filed its US case in the District of Delaware in February 2023. The original complaint alleged copyright infringement, violation of the Digital Millennium Copyright Act (DMCA), and trademark infringement under the Lanham Act. Delaware was chosen because Stability AI, Inc. was incorporated there.

The case moved slowly. Stability AI renewed its motion to dismiss in mid-2024, in part arguing that the Delaware court lacked personal jurisdiction over its UK parent company. With the motion sitting unresolved, Getty took a pragmatic step: it voluntarily dismissed the Delaware case on August 14, 2025, and refiled that same day in the Northern District of California as Getty Images (US), Inc. v. Stability AI, Ltd., case No. 3:25-cv-06891-TLT. The refiled complaint is before Judge Trina L. Thompson and remains active.

Claims in the US Case

The US complaint advances direct copyright infringement, DMCA Section 1202 false copyright management information claims, and Lanham Act trademark claims.

On the DMCA Section 1202 front, Getty alleged that by generating images containing distorted versions of its watermarks — which are a form of copyright management information (CMI) — Stability AI was distributing false CMI in violation of 17 U.S.C. § 1202. The court reviewing the Delaware complaint found that Getty adequately alleged Stability AI’s knowledge that the CMI was false. However, the court dismissed the DMCA claim without prejudice because Getty had not adequately alleged that Stability AI acted with the specific intent required by the statute: the intent to induce, enable, facilitate, or conceal copyright infringement. The door was left open for Getty to replead, and the California filing gives it that opportunity.

The direct copyright infringement claims are the most consequential. Unlike the UK proceeding, the US case will directly confront whether Stability AI’s scraping of Getty’s images for training purposes constitutes infringement and whether the fair use doctrine shields that conduct. Those questions are not resolved in the UK decision.

The Legal Framework: UK and US Diverge Significantly

For businesses trying to assess their exposure, the key point is that UK and US law approach this problem through different doctrinal lenses, and neither has fully answered the question yet.

UK Law: Section 29A CDPA and Its Limits

The UK Copyright, Designs and Patents Act 1988 includes a specific text and data mining (TDM) exception at Section 29A. The exception permits copying works for the purpose of computational analysis where the person conducting the analysis has lawful access to the work, and where the purpose is non-commercial research.

The critical limitation: Section 29A only covers non-commercial research. Commercial AI developers — which includes virtually every company building a product for sale — cannot rely on it. There is no general commercial TDM exception in UK law. This means that, prior to any licensing or judicial resolution, commercial AI training on copyrighted UK works without authorization carries real legal risk.

The UK government spent much of 2025 exploring whether to introduce a broader exception. Following a public consultation that ran from December 2024 through February 2025, the government published its report in March 2026 and announced that it was abandoning its proposed opt-out framework — a model that would have allowed AI training by default unless rights holders affirmatively opted out. The creative industries lobbied strongly against it, and the government concluded it no longer had a preferred policy option. As of mid-2026, UK law defaults to the pre-existing CDPA framework: commercial AI training on copyrighted works without a license remains legally uncertain at best, and potentially unlawful at worst, with the appellate process in Getty v. Stability AI still in motion.

US Law: Fair Use and an Unsettled Landscape

In the United States, there is no equivalent of Section 29A. The applicable doctrine is fair use under 17 U.S.C. § 107, which requires courts to analyze four factors: (1) the purpose and character of the use, including whether it is transformative; (2) the nature of the copyrighted work; (3) the amount and substantiality of what was taken; and (4) the effect of the use on the market for the original work.

Fair use litigation around AI training data is generating conflicting signals. In Bartz v. Anthropic, a court found in June 2025 that using lawfully acquired books to train an AI model was transformative fair use because the purpose — learning statistical relationships in language — differed fundamentally from the original expressive purpose of the books. In Kadrey v. Meta, a court similarly found Meta’s use of books from open-internet datasets transformative for Llama model training. But in Thomson Reuters v. ROSS Intelligence, a court rejected fair use where a company used copyrighted legal headnotes to train a product that competed directly in the market for legal research services.

The US Copyright Office, in its May 2025 Part 3 report on generative AI training, offered a framework rather than a definitive rule. The report found that AI training can be transformative, but that commercial uses enabling competitive market outputs are far less likely to qualify as fair use than non-commercial, non-substitutive research. Critically, the report advanced what it called a “market dilution” theory of harm under the fourth fair use factor: AI systems that flood the market with generated content of the same type as the training works pose a real economic threat to rights holders, even if individual outputs are not direct copies. That analysis is directly relevant to the Getty case, where the argument is that Stable Diffusion enables users to produce images that substitute for licensed Getty photographs.

Trademark Implications: When AI Outputs Display Someone Else’s Logo

The trademark dimension of these cases deserves special attention for any business using AI image tools commercially. The UK High Court’s finding of trademark infringement was based on a specific and concrete injury: outputs that display another company’s registered mark create consumer confusion about source, authorization, or endorsement.

This has broader implications than just the Getty-Stability AI dispute. Businesses using AI image generation tools need to understand that any AI output they publish, distribute, or use commercially becomes their own commercial speech. If that output inadvertently incorporates another company’s trademark — a logo, a watermark, a brand identifier — the business publishing the image is potentially exposed to trademark infringement liability, regardless of whether they knew the AI would produce such an output.

The UK case also establishes that the “we didn’t intend it” defense has limits. Stability AI did not set out to produce Getty-watermarked images, but the court found infringement anyway in the context of historic model versions. A company that uses an AI tool, publishes the output, and later discovers the output contained a third party’s trademark will face a difficult conversation with its counsel about whether it conducted adequate quality review before publication.

There is also a commercial integrity problem. If an AI-generated image your business publishes appears to carry Getty’s or another agency’s watermark, it signals to the audience that your image is an unauthorized version of a licensed work — exactly the opposite of what any legitimate business wants to communicate.

What AI Image Generation Businesses and Enterprise Users Need to Know

For businesses operating in this space, the practical takeaways from these cases fall into two categories: what they mean for AI developers, and what they mean for enterprise customers who use AI image tools in their own operations.

For AI Developers

The UK High Court’s ruling provides some protection for the model weights theory of copyright infringement — the court held that trained weights are not themselves infringing copies. But the primary copyright question (was the scraping itself an infringement?) was never decided because of the UK jurisdictional issue. And Getty has been granted permission to appeal the secondary infringement holding. The ruling is partial, provisional, and appellate scrutiny is coming.

In the US, the direct copyright claims remain live in the California case. US courts have not yet held that scraping copyrighted images for commercial AI training is fair use. In fact, the Thomson Reuters decision and the Copyright Office’s Part 3 report both suggest that commercial, market-competitive uses of copyrighted works face real headwinds under the fourth fair use factor.

AI developers should not read the UK ruling as a green light. It resolved one narrow question under one body of law, with a significant carve-out for the fact that training happened outside the UK. The underlying conduct — taking millions of images without licensing them — remains legally contested on both sides of the Atlantic.

Developers should also pay close attention to trademark hygiene in model outputs. The UK case established liability for trademark-bearing outputs from historic model versions. If your deployed models are capable of generating outputs containing third-party brand marks under realistic prompting, you have a live trademark problem that is not resolved by the copyright discussion.

For Enterprise Customers and Business Users

Companies that use AI image generation tools — whether through API access to Stable Diffusion, Midjourney, DALL-E, Adobe Firefly, or other platforms — are not passive bystanders in this legal landscape. They are the publishers of AI-generated content, and publication is where trademark and copyright liability for outputs most directly attaches.

Most AI image platform terms of service assign output rights to the user while disclaiming any warranty of noninfringement regarding those outputs. Several major platform providers — including Microsoft and Google — have offered some form of IP indemnification to enterprise customers for copyright infringement claims arising from their tools’ outputs, but the scope of those protections varies and frequently excludes uses that violate the platform’s acceptable use policy or that were specifically designed to produce infringing outputs.

Understanding exactly what your platform’s indemnification covers, and what it excludes, is not a task to defer. If you are relying on a vendor’s IP indemnification as your primary shield against output-related liability, you need to read that contract provision carefully.

Due Diligence Checklist for Businesses Using AI Image Tools

The following steps do not eliminate risk in a legally unsettled area, but they represent a defensible, good-faith approach to using AI image generation commercially:

1. Audit your AI image platform’s terms of service. Specifically look for: what rights you receive in generated outputs; whether the platform warrants non-infringement; and what IP indemnification is available for enterprise or business accounts, including which claims it covers and what triggers its exclusions.

2. Implement a human review step before any AI-generated image goes to publication. Review outputs for the presence of logos, watermarks, brand identifiers, or other third-party marks. This is not hypothetical — the Getty watermark phenomenon is documented and reproducible. A simple visual review checkpoint is a basic risk management step.

3. Track your prompts and outputs. For any AI-generated image you publish commercially, retain a record of the prompt used and the platform’s output. This documentation matters if you later need to demonstrate that you acted in good faith and without intent to reproduce another party’s protected content.

4. Avoid prompts designed to elicit style imitation of specific photographers or brands. The closer your prompt is to requesting a specific rights holder’s distinctive style or brand elements, the greater your exposure under both copyright and trademark theories. This is especially true for prompts that explicitly invoke stock photography aesthetics.

5. Understand what “commercial license” actually covers. Many AI image platforms offer a commercial license for outputs. Read what that license covers. It grants you the right to use the output commercially as between you and the platform — it does not indemnify you against third-party copyright or trademark infringement claims arising from the content of that output.

6. Monitor regulatory developments in both the UK and US. The UK Court of Appeal’s decision in Getty v. Stability AI, expected after Getty files its appeal, will be the first appellate-level ruling in any common-law jurisdiction on AI training data copyright. The California case involving the US fair use claims is active and likely to produce rulings within the next year or two. The legal landscape is moving fast.

7. Consider licensing-first approaches where the budget allows. Platforms that train on licensed content — Getty’s own generative AI offerings, Adobe Firefly (trained on Adobe Stock and public domain content), and others — present a materially different risk profile than platforms built on uncleared scrapes of the open web. That distinction may matter significantly if AI training liability rules tighten.

Conclusion

The Getty Images v. Stability AI litigation — in both its UK and US forms — is the most important AI copyright dispute of the current era for businesses that use visual AI tools commercially. The UK High Court’s November 2025 ruling resolved one issue in Stability AI’s favor (model weights are not infringing copies) while leaving the broader copyright training question open and heading to appeal. It simultaneously established that trademark liability is real and present when AI outputs reproduce a rights holder’s brand marks.

The US case, now pending in the Northern District of California, will confront the direct copyright infringement and fair use questions that UK law has not yet answered. The US Copyright Office has already signaled in its Part 3 report that commercial, market-competitive AI training occupies the most vulnerable position in the fair use analysis.

For your business, the practical message is this: the tools your team is using to generate images carry upstream legal uncertainty about how the models were built, and downstream legal exposure for what those models produce and what you do with that output. Neither of those risks is fully resolved by any court ruling to date. Managing them requires understanding your vendor’s contracts, implementing sensible review processes, and staying current as the law continues to develop.

These cases are not just disputes between a large media company and an AI developer. They are the cases that will determine the rules of the road for AI-generated imagery across every industry that uses it. That includes yours.

This post is for informational purposes only and does not constitute legal advice. If you have questions about your specific situation involving AI-generated content, copyright, or trademark, consult qualified legal counsel.