Analysis of Judgment in Getty Images v Stability AI EWHC 2863 (Ch): Legal Failures, Trademark Precedents, and the Shifting Global Litigation Landscape

by Gemini 2.5 Pro, Deep Research. Warning, LLMs may hallucinate! See ‘Take 1’ as an example.

I. Executive Summary of Findings

The High Court judgment in Getty Images (US) Inc & Ors v Stability AI Ltd EWHC 2863 (Ch) delivers a deeply bifurcated outcome that, despite public statements from both parties claiming victory, represents a comprehensive strategic failure for Getty Images on its core copyright claims and a significant, if narrow, victory for Stability AI.¹

While Getty Images secured a partial victory on trademark infringement, the Court’s findings were described by the judge as “both historic and extremely limited in scope”.¹Conversely, Stability AI defeated the entirety of Getty’s copyright case, which was the central, precedential issue at the heart of the litigation.⁵

The key findings of this 201-page judgment are as follows:

1. Primary Copyright Claim (Training) Failed on Jurisdiction: Getty Images was forced to abandon its primary copyright infringement claim after acknowledging there was no evidence that the training of the Stable Diffusion model took place in the United Kingdom.¹ The Court found the training occurred on non-UK cloud servers, rendering the claim territorially inadmissible.¹

2. Secondary Copyright Claim (Importation) Failed on Law: The Court dismissed Getty’s novel secondary infringement claim, delivering a landmark finding for the AI industry. Mrs Justice Joanna Smith DBE ruled that an AI model’s “model weights” are not an “infringing copy” under s.27(3) of the Copyright, Designs and Patents Act 1988 (CDPA). The Court accepted the technical distinction that the model weights are a set of statistical parameters and do not store or contain a copy of the training data.¹

3. Pro-Claimant Finding on “Intangible Articles”: In a crucial finding for future rights holders, the Court rejected Stability’s argument that an “article” under the CDPA must be tangible. The Court affirmed that intangible software, such as an AI model, can be an “article” capable of being “imported” via download.¹

4. Trademark Claims Failed for Newer Models: Getty’s trademark claims against the newer SD XL and v1.6 models failed on the “Threshold Issue.” The Court found Getty had produced “no evidence that one real life user in the UK has generated a watermark” with these models, noting that Getty’s own experiments had failed to do so.¹ This suggests the watermark “bug” had been technically resolved, rendering the claim obsolete.

5. Trademark Tarnishment Claims Failed on Evidence: The claims for reputational damage under s.10(3) TMA were dismissed. The tarnishment claim, which alleged watermarks would appear on pornographic (NSFW) content, failed due to a total lack of real-world evidence of such a co-occurrence.¹ The “unfair advantage” claim failed as it was “wholly unparticularised,” and evidence showed Stability actively tried to remove watermarks as an undesirable flaw.¹

This judgment effectively creates a significant loophole in UK copyright law: AI models trained on infringing data outside the UK can be imported and used in the UK without incurring secondary copyright liability.⁸ This outcome neuters the UK as a jurisdiction for challenging AI training and shifts the definitive legal battleground to the United States, where the substantive question of “Fair Use” remains the central, unresolved issue.⁶

II. Deconstructing the Verdict: A Claim-by-Claim Analysis

The public statements following the judgment were contradictory. Getty Images claimed a “significant win” that confirmed its trademarks were infringed and that Stability AI, not the user, was responsible.³ Stability AI stated it was “pleased with the ruling” which “ultimately resolves the copyright concerns that were the core issue”.²

Both statements are factually correct, but they describe different parts of the case. Getty won a limited battle on trademark law; Stability won the strategic war on copyright law. A definitive analysis requires a claim-by-claim deconstruction of the judgment.

Table 1: Final Verdict and Disposition of Claims in Getty v Stability AI EWHC 2863 (Ch)

Section 1: The Trademark Infringement Claims (Getty’s Pyrrhic Victory)

Getty’s primary public-facing victory was on trademark infringement. This success was, however, severely limited by the Court’s findings on evidence, scope, and intent.

Analysis of the s.10(1) and s.10(2) Win (Older Models)

Getty’s success was twofold. First, the Court found “double identity” infringement under s.10(1) TMA for the iStock marks on model v1.x.¹ Second, it found a likelihood of confusion under s.10(2) TMA for the iStock marks (on v1.x) and the Getty Images marks (on v2.1).¹

The most significant aspect of this finding was the Court’s rejection of Stability’s core defense that the end-user, not Stability, was the party “using” the sign.³ The Court found that Stability’s “active behaviour” and “direct or indirect control” made it liable.¹ This reasoning was built on several pillars:

1. Control over Training: The Court accepted that Stability had “absolute control and/or responsibility” for the training data.¹ Professor Brox, Stability’s own expert, agreed that the reason the models produced watermarks was “because, presumably, Stability failed to remove images bearing Getty Images watermarks from the dataset it trained on”.¹

2. Control over the Model: The Court rejected Stability’s claim that the user has “complete control”.¹ It accepted Professor Farid’s testimony that while the user controls the prompt, they do not control what the model was trained on or its inherent biases. The user “does not have 100% control over what is coming out the other end”.¹

3. Active Provision of Service: Stability was not a passive intermediary. It made the models available via its own platforms (DreamStudio, Developer Platform), which included “post-processing” and “filtering functionality”.¹ This active management and provision of a service constituted “use in the course of trade”.¹

This finding is a critical win for rights holders, as it establishes a clear line of causation from the act of training to liability for the output. The Court affirmed that the model provider, who controls the training, is responsible for the infringing outputs, not the user who prompts it.¹

The “Threshold Issue” Failure (The Evidentiary Collapse for Newer Models)

This victory was immediately and severely curtailed by the Court’s analysis of the “Threshold Issue”: had Getty proven that watermarks* (synthetically generated watermarks) were actually appearing on outputs from the newer SD XL and v1.6 models?

The Court’s answer was a categorical no. This part of the claim collapsed due to a complete failure of evidence.¹⁰

• No Evidence “in the wild”: The Court found “no evidence that one real life user in the UK has generated a watermark using either SD XL or v1.6”.¹ The “Steampunk Musician Image,” Getty’s sole real-world exhibit for SD XL, was dismissed as an image produced during development and “cannot be attributed to the model weights for SD XL after release”.¹

• Getty’s Own Experiments Failed: Crucially, the Court noted that Getty’s ownextensive “Getty Watermark Experiments” had “failed to produce a result for either of these Models”.¹

• “Contrived” Evidence Dismissed: The only evidence of watermarks from these models (the “Donald Glover Image” and “Gabba Images”) was generated by Getty using “verbatim” or “re-worded” prompts (i.e., copying full Getty captions into the prompt box). The Court dismissed these as “contrived and unrepresentative” of real-world use and thus not probative.¹

This failure is highly instructive. It demonstrates that the watermark generation was a technical flaw—a “bug”—in early models. The Court concluded that the lack of evidence “is strongly indicative of the fact that... this issue was resolved using appropriate filters applied to the training data and did not recur in respect of either SD XL or v1.6”.¹ Getty was, in effect, suing over a technical problem that Stability had already fixed, rendering the claim “historic and extremely limited”.¹

Rejection of the s.10(3) Reputation Claims (A Failure of Proof)

Getty’s final trademark claims, alleging damage to its reputation under s.10(3) TMA, were dismissed in their entirety.¹

• Tarnishment (NSFW): Getty’s claim for tarnishment rested on the allegation that its watermarks would appear on “pornography, violent imagery, and propaganda”.¹ This was a central pillar of its public argument. However, the claim failed due to a fundamental lack of evidence.

• The Court accepted expert testimony that the models are capable of producing NSFW content.¹

• The claim failed because Getty could not provide a single real-world example of an image containing both a watermark and NSFW content. The Court stated, “there is not a scrap of evidence that watermarks* have in fact appeared on pornographic images in the real world at any time and in relation to any Model”.¹

• The only evidence provided—the “Miley Cyrus” images—was dismissed as self-generated by Getty’s lawyers “during their experiments” and thus held no evidentiary weight for a tarnishment claim.¹

• This establishes a critical precedent: the theoretical possibility of a brand-damaging co-occurrence is insufficient. A claimant must prove it has actually happened. The Court will not deduce that because A (NSFW content) and B (watermarks) are both possible, A+B has occurred in a way that damages the brand.¹⁰

• Unfair Advantage (Free-Riding): This claim was dismissed just as swiftly.

1. Pleading Failure: The Court found the plea “wholly unparticularised”.¹

2. Lack of Intent: The evidence pointed away from any “free-riding” intent. The Court found “no basis on which I can infer that Stability intended watermarks* to appear”.¹

3. Objective Effect: The “objective effect” was also not one of unfair advantage. Watermarks are “undesirable and make [images] unsuitable for use”.¹ Internal Stability communications showed engineers viewed watermarks as a problem to be fixed, not a feature to exploit. The Court found “Stability sought to take steps to try to filter out watermarks”.¹

III. Incorrect Technical and Legal Assumptions in the Complaint

The judgment reveals that Getty’s core copyright case was built on a series of flawed assumptions, mixing jurisdictional miscalculations with a fundamental technical misunderstanding of how generative AI models function.

Assumption 1: The Jurisdictional Miscalculation (Primary Copyright Claim)

Getty’s primary copyright claim alleged that Stability AI, a UK-registered company, had committed primary infringement under s.17 CDPA by scraping and copying millions of Getty images to train its model.¹ This entire claim was predicated on the assumption that these infringing acts occurred in the United Kingdom.¹²

This assumption was factually incorrect and proved fatal.

• The “Placeless” Cloud: The evidence at trial established that the training and development of Stable Diffusion did not happen on laptops in Stability’s London office. It took place on the “AWS Cluster,” a set of cloud computing resources located outside the United Kingdom.¹

• Territoriality of IP Law: Intellectual property rights are territorial. An act of infringement under the UK’s CDPA must occur in the UK.⁶

• The Abandonment: Faced with this evidence, Getty Images had “no option but to drop the claim”.¹² The Court formally noted the claim was “abandoned”.¹

The judge herself highlighted this jurisdictional failure, stating that Getty “may be able to maintain such a case in the jurisdiction where the Model was in fact trained, but there is no basis for that case in this jurisdiction”.¹ This was a critical strategic error: suing a company in its home jurisdiction is futile if the tortious act occurred in another. For AI development, which relies on “placeless” cloud infrastructure, proving the physical location of data processing is a paramount—and, as Getty discovered, often fatal—evidentiary hurdle.

Assumption 2: The “Model as a Copy” Fallacy (Secondary Copyright Claim)

This was the central, and most ambitious, legal and technical argument in Getty’s case. After its primary claim failed, Getty pivoted to an “ingenious” ⁵ secondary infringement argument under ss. 22 and 23 CDPA.

The argument was:

1. The training process (which happened abroad) would have been an infringement if done in the UK.

2. The result of that process—the AI model’s “model weights”—is therefore an “infringing copy” under s.27(3) CDPA.¹

3. Stability AI “imports” this “infringing copy” into the UK every time a user downloads it, or “distributes” it via its platforms.¹

This entire argument hinged on a single technical consideration: is an AI model’s “model weights” a “copy” of its training data? The Court’s answer, detailed in Section (H) of the judgment, was a definitive no.¹

Mrs Justice Smith’s reasoning was grounded in the technical expert evidence:

1. Models are Not Databases: The Court accepted the expert consensus (including from Stability’s Prof. Brox) that the model weights “do not directly store the pixel values associated with billions of training images”.¹ They are a set of “numerical parameters derived from statistical training”.⁷

2. Storage vs. Statistics: The Court cited the stark difference in file size as dispositive proof: the LAION-5B dataset is 220TB, while the Stable Diffusion v1.4 model weights are 3.44GB. The model is “five orders of magnitude smaller”.¹ It is not a compression or a database of the training images.

3. The Legal Definition of “Copy”: The Court held that for an “article” to be an “infringing copy,” it must be or contain a copy of the work, even if transiently (like the RAM chip in Sony v Ball).¹

4. The Decisive Finding: The Stable Diffusion model “has never consisted of or contained a copy” of Getty’s works.¹ The model weights are “purely the product of the patterns and features which they have learnt”.¹

5. The “Process vs. Product” Distinction: The Court’s final conclusion was that an article (the model) that is merely the product of a process that used infringing copies is not, in itself, an infringing copy.¹

This was the “incorrect technical consideration” at the heart of Getty’s complaint. The legal argument was defeated by the technical reality of how diffusion models function. This finding is a landmark victory for the AI industry, as it insulates the models themselves from being classified as infringing articles under current UK law.⁵

Assumption 3: The “Intangible Article” Precedent (A Key Pro-Claimant Ruling)

While Getty lost the “infringing copy” argument, it won a critical preceding point of law, as highlighted in its own press release.³ Stability had argued that ss. 22 and 23 CDPA could not apply because an “article” must be a tangible object, and an AI model is intangible software.

The Court rejected this argument.¹ In a detailed statutory construction, Mrs Justice Smith found:

• The CDPA does not define “article” as being limited to tangible goods.¹

• The inclusion of s. 27(6), which references works in “electronic form,” suggests that Parliament did envision intangible copies as being “infringing copies”.¹

• The Court concluded that an “article” can be intangible and that the act of downloading software from the internet does constitute “importation” for the purposes of s.22 CDPA.¹

This finding is highly significant for rights holders. While Getty’s claim failed because this specific model was not an “infringing copy,” the Court has affirmed that intangible software can be an “infringing article.” This opens the door for future claims. For example, if an AI model is found to contain copyrighted data (e.g., an LLM that has “memorized” and stores text verbatim), this precedent ensures that ss. 22 and 23 CDPA can be used to block its importation and distribution in the UK.

IV. Consequences for Future AI Litigation

This judgment, while failing to be the “precedential AI and copyright moment” many anticipated ¹³, fundamentally reshapes the legal and strategic landscape for AI litigation, particularly in the UK.

Impact in the United Kingdom

The primary consequence of this ruling is the exposure of a critical gap in UK copyright law, creating what might be termed the “Stable Door” Loophole.

The judgment provides a clear, two-step roadmap for AI companies to operate in the UK while minimizing copyright liability:

1. Train Overseas: Conduct all scraping, copying, and training activities on cloud servers located outside the UK’s jurisdiction. This defeats any claim of primaryinfringement, as Getty discovered.¹

2. Import the Model: The resulting model weights, because they do not contain the original data, are not “infringing copies” under s.27(3) CDPA.¹ Therefore, the model can be freely imported, downloaded, and distributed within the UK without triggering secondary infringement.⁸

This outcome is a “massive damp squib” for rights holders in the UK.⁸ The Court did notrule on the central substantive question: is training an AI model on copyrighted data lawful? That question was avoided on a jurisdictional technicality.⁴

As a result, the battle over AI and copyright in the UK will now shift decisively from the courts to Parliament. The CDPA, enacted in 1988, is ill-equipped for generative AI. This judgment places enormous pressure on the UK government to legislate, as advocated by groups like the Publishers Association.¹³ The key policy debates will now concern mandatory transparency requirements for training data and potential amendments to the UK’s Text and Data Mining (TDM) exception.⁵

Impact in the United States (The Real Battleground)

The Getty UK judgment’s primary global consequence is to reinforce that the only place the core training debate can be effectively litigated is in the jurisdiction where the training occurred—which, for most major models, is the United States.⁴

This ruling highlights the stark contrast between the UK and US legal frameworks:

• The UK case failed on procedure (jurisdiction) and technical definition (what is a “copy”?).

• The US cases (e.g., Bartz v. Anthropic, Kadrey v. Meta, Thomson Reuters v. Ross) are being fought on the substantive affirmative defense of “Fair Use” under 17 U.S.C. § 107.¹⁵

In the US, the litigation is focused on two key factors of the Fair Use test:

1. Factor 1 (Transformative Use): AI companies argue that training is “transformative” because the model is a new tool, not a substitute for the original works. This argument has found initial success in US district courts, such as in Bartz v. Anthropicand Kadrey v. Meta.¹

2. Factor 4 (Market Harm): This is the rights holders’ strongest counter-argument. Recent rulings, such as in Thomson Reuters v. Ross ²⁰ and the Kadrey decision, suggest that demonstrating concrete market substitution or “market dilution” is the key to defeating a fair use defense.¹⁵

3. Data Provenance: A critical factor emerging in US law is the source of the training data. The Bartz court, while finding training on legally purchased books could be fair use, was deeply skeptical of training on pirated libraries, calling it “inherently, irredeemably infringing”.¹⁵

The UK judgment, by closing the door on secondary infringement claims, inadvertently channels all future high-stakes copyright litigation to the US to be fought on these substantive Fair Use grounds.

V. Strategic Recommendations for Rights Holders

This judgment necessitates a fundamental reassessment of legal and commercial strategy for all rights holders.

The New Litigation Playbook (Post-Getty)

1. Prioritize Jurisdiction Above All Else: Cease all attempts to litigate AI training in jurisdictions where the training did not physically occur. The Getty UK case proves this is a fatal and costly error.⁴ All resources must be consolidated and focused on the locus of training, which is, for the foreseeable future, the United States.

2. Abandon the “Model is a Copy” Argument: The UK Court’s technical and legal rejection of this theory is comprehensive, well-reasoned, and based on expert consensus.¹ It is highly persuasive. Continuing to advance this argument in other common law jurisdictions is a high-risk, low-reward strategy.

3. Pivot US Strategy to Market Harm (Factor 4): The legal focus in the US mustpivot from “you copied” (which may be deemed “transformative”) to “your output destroys our market.” Rights holders must commission economic experts and gather concrete evidence of market dilution and indirect substitution.¹⁵ As one US judge noted, this may be the path for plaintiffs to “decisively win the fourth factor—and thus win the fair use question overall”.¹⁵

4. Weaponize Data Provenance: Discovery in US cases must aggressively target the source of the training data. The Bartz decision provides a powerful weapon: evidence that an AI company knowingly used pirated datasets (e.g., “Books3”) may be the most direct path to defeating a Fair Use defense, regardless of the “transformative” nature of the model.¹⁵

Legislative and Lobbying Priorities

1. UK Legislative Reform: The Getty judgment is a powerful lobbying tool. Rights holders must immediately petition the UK government to amend the CDPA to close the s.27 “infringing copy” loophole. The law must be updated to state that an article (like an AI model) created by an infringing process is itself an infringing article, regardless of whether it “contains” a copy.

2. Global Transparency Mandates: The greatest barrier Getty faced was proving its works were used and where they were used.¹³ Rights holders must form a unified front to lobby globally (UK, US, EU) for mandatory “AI nutrition label” laws. These laws must require AI developers to disclose their training data sources as a precondition of market access. This solves the evidentiary problem at the heart of all future litigation.

Commercial and Defensive Strategies

1. Leverage Trademark Law (At the Output): The one clear win for Getty was that an AI provider is liable for trademark-infringing outputs.¹ This remains a viable, albeit limited, tool. It places the onus (and cost) of filtering on the provider, not the user, which is a valuable precedent.³

2. Negotiate for Technical Filters: The judgment shows that technical “bugs” like watermark generation are fixable.¹ Rights holders should use this as leverage in all licensing negotiations, demanding robust, state-of-the-art output filtering against registered trademarks, known licensed content, and brand-damaging (e.g., NSFW) associations as a non-negotiable condition.

3. Re-evaluate Collective Licensing: The High Court previously rejected Getty’s attempt to bring a representative action on behalf of all its contributors.¹ This judgment confirms that the legal and technical hurdles of individual litigation are prohibitively complex and expensive.¹³ This outcome paradoxically strengthens the case for collective licensing bodies, which can negotiate on behalf of thousands of creators from a position of unified strength, rather than relying on fragmented, high-risk litigation.

Works cited

1. Getty-Images-v-Stability-AI

2. Stability AI largely wins UK court battle against Getty Images over copyright and trademark, accessed November 5, 2025, https://apnews.com/article/getty-stability-ai-image-copyright-trademark-fa2c561a33c7b6714a7657255a3fbdf1

3. Getty Images issues statement on ruling in Stability AI UK litigation ..., accessed November 5, 2025, https://newsroom.gettyimages.com/en/getty-images/getty-images-issues-statement-on-ruling-in-stability-ai-uk-litigation

4. Getty vs Stability AI ruling is ‘damp squib’ that fails to create ..., accessed November 5, 2025, https://pressgazette.co.uk/media_law/getty-vs-stability-ai-copyright-ruling-uk/

5. Getty Images v Stability AI - Getty unsuccessful in secondary ..., accessed November 5, 2025, https://cms-lawnow.com/en/ealerts/2025/11/getty-images-v-stability-ai-getty-unsuccessful-in-secondary-copyright-infringement-claim-re-stable-diffusion

6. UK court mostly sides with Stability AI in AI copyright lawsuit ..., accessed November 5, 2025, https://siliconangle.com/2025/11/04/uk-court-mostly-sides-stability-ai-ai-copyright-lawsuit/

7. The UK’s First Copyright vs AI Decision: Key Takeaways on a Win for ..., accessed November 5, 2025, https://www.sidley.com/en/insights/newsupdates/2025/11/uk-first-copyright-vs-ai-decision-key-takeaways-on-a-win-for-the-ai-industry

8. Getty loses UK copyright battle against Stability AI • The Register, accessed November 5, 2025, https://www.theregister.com/2025/11/04/uk_court_getty_stability_ai/

9. AI firm wins high court ruling after photo agency’s copyright claim, accessed November 5, 2025, https://www.theguardian.com/media/2025/nov/04/stabilty-ai-high-court-getty-images-copyright

10. Getty Images v Stability AI - Courts and Tribunals Judiciary, accessed November 5, 2025, https://www.judiciary.uk/wp-content/uploads/2025/11/Getty-Images-v-Stability-AI.pdf

11. Getty Images v Stability AI: where are we after the trial - trade marks? - Taylor Wessing, accessed November 5, 2025, https://www.taylorwessing.com/en/insights-and-events/insights/2025/09/getty-v-stability-where-are-we-after-the-trial---trade-marks

12. Getty Images v Stability AI: where are we after the trial - copyright?, accessed November 5, 2025, https://www.taylorwessing.com/en/insights-and-events/insights/2025/07/getty-v-stability

13. Statement following the Getty Images v Stability AI judgment ..., accessed November 5, 2025, https://www.publishers.org.uk/statement-following-the-getty-images-v-stability-ai-judgement/

14. How a US AI ruling could influence UK copyright policy - Pinsent Masons, accessed November 5, 2025, https://www.pinsentmasons.com/out-law/analysis/how-a-us-ai-ruling-could-influence-uk-copyright-policy

15. A Tale of Three Cases: How Fair Use Is Playing Out in AI Copyright ..., accessed November 5, 2025, https://www.ropesgray.com/en/insights/alerts/2025/07/a-tale-of-three-cases-how-fair-use-is-playing-out-in-ai-copyright-lawsuits

16. AI Cases in the UK and the US: What Could They Mean for AI ..., accessed November 5, 2025, https://www.tilleke.com/insights/ai-cases-in-the-uk-and-the-us-what-could-they-mean-for-ai-globally/

17. The UK Getty Trial: Key Takeaways on the AI/Copyright Case | Insights | Jones Day, accessed November 5, 2025, https://www.jonesday.com/en/insights/2025/07/the-uk-getty-trial-key-takeaways-on-the-aicopyright-case

18. Artificial Intelligence: Latest Government Regulations and Copyright Litigation Updates, accessed November 5, 2025, https://www.foster.com/newsroom-alerts-artificial-intelligence-latest-government-regulations-and-copyright-litigation-updates

19. Mid-Year Review: AI Copyright Case Developments in 2025, accessed November 5, 2025, https://copyrightalliance.org/ai-copyright-case-developments-2025/

20. Court Grants Summary Judgment in AI Copyright Clash, Rejecting “Fair Use” - Jones Day, accessed November 5, 2025, https://www.jonesday.com/en/insights/2025/02/no-fair-use-in-ai-copyright-clash

21. From Input to Impact: The Market Harm Standard Emerging in AI Fair Use | Alerts and Articles | Insights | Ballard Spahr, accessed November 5, 2025, https://www.ballardspahr.com/insights/alerts-and-articles/2025/07/from-input-to-impact-the-market-harm-standard-emerging-in-ai-fair-use