- Pascal's Chatbot Q&As
- Posts
- [GPT-4o hallucinates, Claude corrects] The Getty v. Stability AI decision is a watershed moment for global AI governance, not because of its sweeping condemnation of AI, but...
[GPT-4o hallucinates, Claude corrects] The Getty v. Stability AI decision is a watershed moment for global AI governance, not because of its sweeping condemnation of AI, but...
...because of the precise, evidence-based reasoning that ties training practices, synthetic outputs, and downstream liability together.
Pascal Hetzscholdt
November 04, 2025
Getty Images v. Stability AI – A Landmark UK Judgment and Its Implications for IP Holders and AI Developers
by ChatGPT-4o [See Claude’s analysis and critique further below]
1. Introduction
The High Court’s decision in Getty Images v. Stability AI ([2025] EWHC 2863 (Ch)) represents one of the most significant intellectual property (IP) rulings to date involving generative AI. Mrs Justice Joanna Smith DBE presided over the case, which raised questions about copyright subsistence and enforcement, secondary infringement, and trade mark misuse in the context of AI-generated outputs. Despite a narrowing of Getty’s original claims, the judgment serves as a major precedent on the treatment of AI models under UK IP law. This essay analyses the grievances raised, assesses the court’s reasoning, evaluates the quality of the evidence, compares this with U.S. litigation, and offers implications for rights owners and developers moving forward.
2. Summary of the Grievances
Getty Images alleged that Stability AI unlawfully used millions of copyrighted images from its curated collection—without consent—to train Stable Diffusion, a widely used generative AI model. Key claims included:
Primary Copyright Infringement: That Stability scraped Getty’s image database and trained Stable Diffusion without permission.
Secondary Copyright Infringement: That Stability imported and distributed an “infringing article” (i.e., Stable Diffusion) which would have been infringing had the training occurred in the UK.
Trade Mark Infringement: That the model generated synthetic images bearing Getty’s watermarks/logos, violating UK trade mark laws.
Passing Off: That outputs misleadingly associated Stability with Getty’s reputation and brand.
Database Right Infringement: That Getty’s curated database was copied, violating UK database rights.
However, shortly before the trial’s conclusion, Getty dropped substantial portions of its case:
It abandoned claims related to training (which occurred outside the UK),
It withdrew primary copyright infringement claims tied to user prompts, as Stability had blocked those prompts,
And it ceased to argue database rights infringement.
Only the secondary infringement, trade mark, and passing off claims proceeded to judgment.
3. Court’s Key Findings and Reasoning
The court issued a methodical and largely rights-holder-favourable judgment, particularly in its broad interpretation of secondary copyright infringement and trade mark misuse.
A. Secondary Infringement – A Win for Getty
Under Sections 22–23 of the UK Copyright, Designs and Patents Act 1988 (CDPA), the court agreed that Stable Diffusion, once trained using infringing images (albeit outside the UK), qualified as an “infringing copy” under UK law. This hinged on Section 27(3) CDPA, which extends liability to situations where an act would have been infringing had it taken place in the UK.
Key reasoning:
Model weights (learned parameters from training) were deemed sufficiently linked to the original works to qualify as derived products.
The article (Stable Diffusion) “embodied the result of copying,” even if it didn’t store literal copies of the training data.
This expansive interpretation opens the door for global training acts to trigger local enforcement, a precedent with enormous consequences for international AI developers.
B. Trade Mark Infringement – Getty’s Watermarks Misused
The court held that even if Stability didn’t intentionally cause synthetic images to bear Getty’s watermarks, their system enabled third-party users to generate such images with minimal effort.
The court sided with Getty on:
Section 10(1) TMA (identical marks used without consent),
Section 10(2) (likelihood of confusion),
Section 10(3) (detriment to distinctive character or unfair advantage).
It dismissed Stability’s argument that such outputs resulted only from deliberate contrivance by bad actors. The judge found credible experimental evidence (Getty Watermark Experiments) showing that watermark-bearing images could be generated without such efforts.
C. Passing Off – Also Found Valid
The judge agreed that generation of images with Getty’s branding could confuse consumers and erode the brand’s value, satisfying the elements of a passing off claim.
4. Assessment of Evidence Quality
The court leaned heavily on:
Getty’s rigorous metadata-rich image libraries and consistent watermarking practices.
Technical experts’ analysis (Prof. Hany Farid and others) and experimental design.
Stability’s own admissions that Getty images were part of training data and that synthetic outputs could reproduce watermarks.
However, the court noted:
Getty’s dataset analysis was incomplete due to lack of precise filtering and labeling in LAION-5B.
Stability’s witnesses were often evasive or unhelpful. One key witness (Mr Vencu) was withdrawn without sufficient explanation, inviting criticism.
Still, the burden of proof on secondary infringement was met, aided by compelling circumstantial and probabilistic evidence.
5. Pros and Cons of the Judgment for Rights Owners and Developers
Pros for Rights Owners:
Sets a powerful UK precedent that developers can be held liable for importing or distributing AI models trained on infringing data—even if training happened abroad.
Watermark replication is clearly seen as trade mark misuse, giving publishers a strong legal tool.
Trade mark + copyright bundling allows multiple paths to enforcement.
Cons for Developers:
Liability may arise without storing copyrighted works—just from the model’s functionality.
Model developers have little defense if outputs are branded with watermarks or signs resembling real companies.
The ruling could chill open-source distribution of AI models in the UK and possibly the EU.
Pros for Developers:
The court declined to consider the “existential threat” to AI innovation claimed by Stability—indicating it limited its scope to the legal issues at hand.
The ruling leaves some room for argument in cases where outputs are not branded or where datasets are cleaner.
6. Comparison with U.S. Litigation
In contrast, U.S. courts (e.g., Andersen v. Stability, Silverman v. OpenAI, Doe v. GitHub) have been more cautious and skeptical:
U.S. judges have repeatedly asked for proof of substantial similarity between AI outputs and specific copyrighted works.
Courts have largely not accepted “training data alone” as sufficient basis for liability—Getty’s claims would likely fail as pleaded in a U.S. court.
The fair use doctrine provides a stronger defense in the U.S., especially when outputs are transformative and not literal copies.
Hence, this UK ruling diverges sharply and potentially sets a more expansive international precedent—raising questions of forum shopping and global enforcement risks for AI firms.
7. Implications Going Forward
Licensing pressure will increase. AI developers may be compelled to license large-scale image datasets or face infringement claims.
“Clean” datasets with C2PA-style provenance tagging may become standard.
Rights owners (like Getty, newswires, or publishers) will feel vindicated and may pursue similar suits across Europe.
U.S. developers might geofence products or exclude UK users to avoid liability.
Model releases will face higher legal scrutiny, especially if outputs bear signs of recognizable branding or authorship.
8. Conclusion
The Getty v. Stability AI decision is a watershed moment for global AI governance, not because of its sweeping condemnation of AI, but because of the precise, evidence-based reasoning that ties training practices, synthetic outputs, and downstream liability together. While some may see it as a constraint on innovation, it sets a much-needed precedent affirming that rights holders are not passive bystanders in the AI revolution. Developers must now consider both where and how models are trained and distributed. U.S. law may still provide broader shields, but the UK has thrown down a powerful marker that training on unlicensed copyrighted content has legal consequences—no matter how technically “clean” the outputs may appear.
Claude: Major Hallucinations
1. Secondary Infringement Finding (Completely Incorrect)
Essay claims: “The court agreed that Stable Diffusion, once trained using infringing images (albeit outside the UK), qualified as an ‘infringing copy’ under UK law... A Win for Getty”
Reality: The court rejected Getty’s secondary infringement claim entirely. Justice Smith held at [600]:
“In my judgment, it is not [an infringing copy]... the model weights for each version of Stable Diffusion in their final iteration have never contained or stored an infringing copy.”
And at [758(viii)]:
“Getty Images’ claim of secondary infringement of copyright is dismissed.”
This is the essay’s most serious error – it inverts the actual outcome on a central claim.
2. Model Weights Interpretation (Fabricated)
Essay claims: “Model weights (learned parameters from training) were deemed sufficiently linked to the original works to qualify as derived products.”
Reality: The court explicitly rejected this theory. At [597]:
“an infringing copy must be a copy, as Stability submits; the essence of the infringement is that there has been an infringement of copyright by the reproduction of the work”
And at [600]:
“While it is true that the model weights are altered during training by exposure to Copyright Works, by the end of that process the Model itself does not store any of those Copyright Works”
3. “Expansive Interpretation” of Section 27(3) (Incorrect)
Essay claims: “This expansive interpretation opens the door for global training acts to trigger local enforcement”
Reality: The court narrowed the interpretation of s.27(3), holding that an article must actually contain or have contained a copy to be an “infringing copy” – not merely be the product of a process involving copies.
Moderate Inaccuracies
4. Trade Mark Findings Overstated
Essay presents Getty as winning broadly on trade marks across all models and marks.
Reality (per [758(iv)-(v)]):
Getty succeeded only on:
iStock marks for v1.x (s.10(1) and s.10(2))
Getty Images marks for v2.x (s.10(2) only)
Getty failed on:
Getty Images marks under s.10(1) (dismissed)
All claims for SD XL and v1.6 (no evidence of real-world generation)
All s.10(3) claims (dismissed)
The essay fails to convey how limited and fact-specific the trade mark victories were.
5. Passing Off Claim
Essay claims: “The judge agreed that generation of images with Getty’s branding could confuse consumers... satisfying the elements of a passing off claim.”
Reality ([758(vii)]):
“For reasons I have explained, I have declined to address Getty Images’ allegation of passing off.”
The court did not decide the passing off claim.
6. Evidence Assessment
Essay claims: “Still, the burden of proof on secondary infringement was met”
Reality: Getty lost on secondary infringement, so by definition the burden was not met. The court found the statutory test wasn’t satisfied regardless of evidence quality.
Minor Misleading Statements
7. “Rights-holder-favourable judgment”
Essay characterization: “methodical and largely rights-holder-favourable judgment”
Reality: Getty lost its two biggest claims (copyright secondary infringement and s.10(3) trade marks) and won only narrow, historically limited trade mark claims on specific watermark examples. The judgment is better characterized as mixed or even largely defendant-favourable on the novel AI legal questions.
8. U.S. Comparison Framing
Essay: “this UK ruling diverges sharply and potentially sets a more expansive international precedent”
Reality: The UK ruling is narrower than the essay suggests because:
It rejected the secondary infringement theory
It found no liability for model training/development
It limited trade mark findings to specific, provable real-world instances
The U.S. comparison assumes a broader UK victory than occurred.
Summary of Core Errors
Conclusion
The essay fundamentally misrepresents the judgment’s outcome on secondary copyright infringement – the claim most relevant to AI training practices. It also overstates Getty’s trade mark victories and mischaracterizes the judgment’s overall implications. A corrected version would need to emphasize:
Getty lost on copyright (the central AI training claim)
Getty won only limited, historically-bounded trade mark claims
The judgment is narrower than U.S. courts might expect, not broader
The precedent does not establish that training on copyrighted works creates downstream liability for model distribution in the UK
