Home » Blog » Getty Images (US) Inc and others v Stability AI Ltd

Getty Images (US) Inc and others v Stability AI Ltd

By /

Authored By: Nithishwaran P S

Brunel University London

  1. Case citation and basic information
  • Full Case Name: Getty Images (US) Inc and others v Stability AI Ltd.[1]
  • Citation (neutral): EWHC 2863 (Ch).[2]
  • Court: High Court of Justice, Business and Property Courts of England and Wales, Intellectual Property List (Chancery Division).[3]
  • Date of Decision: 4 November 2025.[4]
  • Bench Composition: Joanna Smith J (single judge).[5]
  1. 2. Introduction

This case concerns whether UK copyright and trade mark law apply to generative AI models trained on large datasets of images without permission, specifically the Stable Diffusion image generator developed by Stability AI.[6] Getty Images, a major stock image provider, alleged that Stability had used millions of its photographs and related content to train Stable Diffusion and that the resulting model and outputs infringed copyright, database rights and registered trademarks.[7] The decision is significant as the first detailed judicial analysis of whether AI model weights constitute “infringing copies” under the Copyright, Designs and Patents Act 1988 (CDPA).[8] While Getty achieved limited success on trademark claims concerning watermark reproduction, its core copyright arguments failed, providing important clarity for AI developers while leaving key questions about training data legality unresolved.[9]

  1. Facts of the case

Getty operates a global stock photography and licensing business, maintaining extensive libraries of images and videos and licensing them to media, advertisers and other professional users.[10] Many of these works are protected by copyright and associated database rights, and the “Getty Images” word and device marks are protected by registered trademarks in the UK and other jurisdictions.[11] Stability AI Ltd is the UK company associated with Stable Diffusion, a text‑to‑image generative AI model which produces images in response to text prompts by users.[12]

Getty alleged that Stability, directly or via collaborators, had used a very large corpus of online images that included millions of Getty works, as part of a dataset (often described in the commentary as LAION‑5B) to train Stable Diffusion.[13] Training involved repeated processing of images and their captions to adjust the model’s “weights” – numerical parameters representing learned correlations.[14] Getty argued that this process necessarily copied its images and that the trained model weights embodied copies or substantial parts of those works.[15]

Getty advanced three main categories of claim.[16] First, it alleged primary copyright and database right infringement arising from the training and development of Stable Diffusion using Getty images, as well as from allegedly infringing outputs.[17] Second, it claimed secondary copyright infringement under sections 22–23 CDPA, contending that the model made available in the UK was an “article” which was itself an infringing copy.[18] Third, it alleged trade mark infringement and passing off based on AI‑generated images which reproduced the Getty watermark, arguing that these outputs took unfair advantage of, and were detrimental to, the distinctiveness and reputation of its marks.[19]

During the proceedings, Getty conceded that it lacked evidence that the actual training and development of Stable Diffusion occurred within the UK, and consequently abandoned its primary copyright and database claims tied to UK‑territorial acts.[20] Getty likewise narrowed its claims concerning specific user-generated outputs, leaving the High Court to decide the secondary copyright claim and the trade mark and passing‑off allegations relating to watermark-bearing outputs.[21]

  1. Legal issues

Issue 1: Whether the Stable Diffusion model (particularly its weights) constituted an “infringing copy” of Getty’s copyrighted works for the purposes of secondary copyright infringement under sections 22–23 CDPA 1988.[22]

Issue 2: Whether making Stable Diffusion available in the UK – by hosting, distribution or provision of access – amounted to dealing with such “infringing copies” in the course of business, thereby engaging secondary liability under the CDPA.[23]

Issue 3: Whether the generation of images carrying the Getty watermark by certain versions of Stable Diffusion infringed Getty’s registered trademarks under sections 10(1), 10(2) or 10(3) of the Trade Marks Act 1994.[24]

Issue 4: Who bears legal responsibility for watermark‑bearing outputs – the model developer (Stability) or end‑users who submit prompts – and whether any passing‑off liability arises from those outputs.[25]

  1. Arguments of the parties

Getty (claimants)

Getty argued that the training process necessarily entailed copying its images and that the resulting Stable Diffusion model, expressed in its weights, embodied those works in a compressed form.[26] On this basis, Getty submitted that the model weights were “articles” which were infringing copies under section 27 CDPA, so that Stability’s making the model available to UK users amounted to secondary infringement under sections 22–23.[27] Getty stressed that the economic value of its library lay precisely in the ability to license large‑scale uses such as training, and that allowing unlicensed use would undermine this business.[28]

On trademarks, Getty contended that images generated by certain early Stable Diffusion versions reproduced the GETTY IMAGES watermark, leading to a likelihood of confusion and taking unfair advantage of the mark’s reputation.[29] It argued that Stability’s choice to include watermark‑bearing images in training data, and its failure to implement adequate technical safeguards, made it directly responsible for these infringing signs.[30]

Stability AI (defendant)

Stability denied that the model weights contained any copies of Getty’s works, emphasising expert evidence that weights store statistical parameters rather than pixel‑based reproductions of training images.[31] It argued that an “infringing copy” under the CDPA requires an article which actually reproduces the work or a substantial part, which was not the case here.[32] Stability further contended that training had taken place outside the UK and that the abandoned primary claims could not be repackaged as secondary liability.[33]

Regarding trade marks, Stability argued that any watermark‑bearing outputs were rare artefacts generated by user prompts, and that it should not be held strictly liable for every image produced by users of an open‑ended generative model.[34] It also maintained that there was no real likelihood of confusion or damage to the Getty brand from a small number of experimental outputs.[35]

  1. Court’s reasoning and analysis

(a) Copyright and “infringing copy”

The judge began by examining the statutory framework for secondary infringement, which requires, among other elements, the dealing in the course of business with an “article” which is an “infringing copy” of a copyright work.[36] Drawing on prior case law about temporary digital copies (e.g. RAM copies), the court accepted that non‑permanent digital embodiments can be copies, but only where they themselves contain the work or a substantial part of it.[37]

On the expert evidence, the court found that the Stable Diffusion weights did not store or reproduce the Getty images, but instead represented learned parameters that enable generation of new images without retaining copies of specific training works.[38] The judge rejected Getty’s characterisation of the model as a compressed copy of the entire training set, holding that this “over‑extends” the concept of a copy beyond the statutory language.[39]

Because the weights did not themselves embody Getty’s works, they could not be “infringing copies” under section 27 CDPA.[40] Without an infringing copy, the secondary infringement provisions of sections 22–23 were not engaged, and the claim necessarily failed.[41] The judge noted that Getty’s abandonment of its primary training‑based claims meant the court was not deciding whether UK‑based training on protected images would be lawful; that issue remains open.[42]

(b) Trademarks and responsibility for outputs

Turning to trade marks, the court accepted evidence that a very small number of images generated by early model versions reproduced the Getty watermark to a high degree of similarity.[43] For those outputs, the judge held that there was trade mark infringement under sections 10(1) and 10(2) TMA 1994, given the use of an identical or similar sign for identical or similar services and the resulting likelihood of confusion.[44]

However, the court found that the overall scope of infringement was “historic and extremely limited” in light of the tiny number of proven instances and improvements in later versions of the model.[45] The broader claim under section 10(3) – based on reputation and unfair advantage or detriment – was rejected as insufficiently substantiated on the facts.[46]

On responsibility, the judge held that Stability, not end‑users, bore primary liability for these limited infringements, emphasising that developers exercise control over training data, filters and safeguards and cannot fully shift responsibility to users when outputs reproduce protected signs.[47] The passing‑off claim was treated as adding little beyond the trade mark analysis and did not materially expand Stability’s liability.[48]

  1. Judgment and ratio decidendi

The High Court dismissed Getty’s secondary copyright and database right claims in their entirety, holding that Stable Diffusion’s model weights are not “infringing copies” of Getty’s images within the meaning of the CDPA.[49] The court granted only very limited relief for trade mark infringement in relation to a small number of historical watermarks‑bearing outputs, rejecting broader claims under section 10(3) TMA and declining to extend liability via passing off.[50]

The ratio decidendi on copyright is that a trained AI model’s parameters, which encode statistical relationships but do not store or reproduce the underlying works, do not constitute an “infringing copy” of those works for the purposes of secondary infringement.[51] On trademarks, the binding principle is more limited: where a generative model produces outputs reproducing a registered mark because of the developer’s choices about data and safeguards, the developer can be directly liable, but isolated, historic instances may justify only narrow relief.[52]

  1. Critical analysis (your original contribution)

Getty v Stability AI offers considerable comfort to AI developers by rejecting the notion that a foundation model is, by its nature, an infringing copy of its training set, but it does so by focusing closely on the technical characteristics of model weights and the statutory wording of “copy”.[53] This approach arguably keeps copyright doctrine within traditional boundaries, yet it leaves unresolved the more pressing normative question: whether large-scale, unlicensed training on protected images should itself be lawful when conducted within the UK.[54]

The judgment has been criticised for creating a potential enforcement gap: if training takes place outside the jurisdiction and the resulting model is not itself a copy, UK rightsholders may struggle to challenge the exploitation of their works in global AI systems.[55] At the same time, the trademark reasoning illustrates a willingness to attribute responsibility to developers for problematic outputs, which may push industry practice towards more cautious dataset curation and watermark filtering.[56]

From a policy perspective, the decision highlights the limits of adapting existing IP concepts to radically new technologies.[57] Legislatures may need to clarify whether and on what terms training on protected content is allowed, rather than leaving the issue to incremental judicial development through cases constrained by territorial and evidential difficulties.[58]

  1. Conclusion

Getty v Stability AI is a landmark in the interaction between AI and IP, but it is not the final word.[59] It establishes that, under current UK law, a generative model’s parameters are not in themselves infringing copies, thus defeating secondary copyright claims of the kind advanced by Getty.[60] At the same time, the narrow finding of trademark infringement for watermark-bearing outputs shows that developers cannot ignore the risk of replicating brand identifiers in training and deployment.[61]

The key takeaway is that existing IP frameworks can address some AI-related harms, but major questions—especially concerning the lawfulness of training itself—remain open and are likely to be tested further in future litigation and possible legislative reform.[62]

  1. BIBLIOGRAPHY

[1] Getty Images (US) Inc and others v Stability AI Ltd EWHC 28.

[2] ibid.

[3] ibid.

[4] ibid.

[5] ibid.

[6] Getty Images (US) Inc and others v Stability AI Ltd EWHC 2863 (Ch).

[7] Latham & Watkins, ‘Getty Images v Stability AI: English High Court rejects secondary copyright claim’ (Latham & Watkins, 3 November 2025) <https://www.lw.com/en/insights/getty-images-v-stability-ai-english-high-court-rejects-secondary-copyright-claim> accessed 10 February 2026.

[8] William Fry, ‘Getty Images v Stability AI – The Most Important AI Legal Decision to Date’ (William Fry, 5 November 2025) <https://www.williamfry.com/knowledge/getty-images-v-stability-ai-the-most-important-ai-legal-decision-to-date/> accessed 10 February 2026.

[9] Mishcon de Reya, ‘Getty Images v Stability AI: Unpacking the High Court’s judgment’ (Mishcon de Reya, 5 November 2025) <https://www.mishcon.com/news/getty-images-v-stability-ai-unpacking-the-high-courts-judgment> accessed 10 February 2026.

[10] Associated Press, ‘Stability AI largely wins UK court battle against Getty Images over copyright and trademark’ (AP News, 4 November 2025) <https://apnews.com/article/getty-stability-ai-image-copyright-trademark-fa2c561a33c7b6714a7657255a3fbdf1> accessed 10 February 2026.

[11] Getty Images (US) Inc and others v Stability AI Ltd EWHC 2863 (Ch) -.

[12] Latham & Watkins (n 7).

[13] Keystone Law, ‘Getty Images v Stability AI: the key intellectual property issues’ (Keystone Law, 9 December 2025) <https://www.keystonelaw.com/keynotes/getty-images-v-stability-ai-the-key-intellectual-property-issues> accessed 10 February 2026.

[14] Clayton Utz, ‘Getty Images v Stability AI: implications for intellectual property and data protection’ (Clayton Utz, 25 November 2025). <https://www.claytonutz.com/insights/2025/november/getty-images-v-stability-ai-implications-for-intellectual-property-data-protection-and-ai-regulation> accessed 10 February 2026.

[15] Getty Images (n 1) -.

[16] William Fry (n 8).

[17] Getty Images (n 1) .

[18] Copyright, Designs and Patents Act 1988, ss 22-23.

[19] Trade Marks Act 1994, ss 10(1)-(3).

[20] Latham & Watkins (n 7).

[21] Mishcon de Reya (n 9).

[22] Clayton Utz (n 14).

[23] Mishcon de Reya (n 9).

[24] ABC News, ‘Stability AI largely wins court battle against Getty Images over copyright, trademark’ (ABC News, 3 November 2025) <https://abcnews.go.com/Technology/wireStory/stability-ai-largely-wins-uk-court-battle-getty-127164244> accessed 10 February 2026.

[25] Osborne Clarke, ‘Stability AI generates big win in English court’s landmark first judgment on AI image generators’ (Osborne Clarke, 5 November 2025) <https://www.osborneclarke.com/insights/getty-v-stability-ai-stability-ai-generates-big-win-english-courts-landmark-first-judgment> accessed 10 February 2026.

[26] Associated Press, ‘Stability AI largely wins UK court battle against Getty Images over copyright and trademark’ (AP News, 4 November 2025) https://apnews.com/article/getty-stability-ai-image-copyright-trademark-fa2c561a33c7b6714a7657255a3fbdf1 accessed 10 February 2026.

[27] Latham & Watkins (n 7).

[28] Getty Images (US) Inc and others v Stability AI Ltd EWHC 2863 (Ch) – https://www.judiciary.uk/wp-content/uploads/2025/11/Getty-Images-v-Stability-AI.pdf accessed 10 February 2026

[29] ABC News, ‘Stability AI largely wins court battle against Getty Images over copyright, trademark’ (ABC News, 3 November 2025) https://abcnews.go.com/Technology/wireStory/stability-ai-largely-wins-uk-court-battle-getty-127164244 accessed 10 February 2026.

[30] ibid.

[31] Keystone Law, ‘Getty Images v Stability AI: the key intellectual property issues’ (Keystone Law, 9 December 2025) https://www.keystonelaw.com/keynotes/getty-images-v-stability-ai-the-key-intellectual-property-issues accessed 10 February 2026.

[32] Clayton Utz (n 14).

[33]Latham & Watkins (n 7).

[34] Osborne Clarke, ‘Stability AI generates big win in English court’s landmark first judgment on AI image generators’ (Osborne Clarke, 5 November 2025) https://www.osborneclarke.com/insights/getty-v-stability-ai-stability-ai-generates-big-win-english-courts-landmark-first-judgment accessed 10 February 2026.

[35] ibid.

[36] William Fry, ‘Getty Images v Stability AI – The Most Important AI Legal Decision to Date’ (William Fry, 5 November 2025) https://www.williamfry.com/knowledge/getty-images-v-stability-ai-the-most-important-ai-legal-decision-to-date/ accessed 10 February 2026.

[37]  Getty Images (n 1) -.

[38] Keystone Law (n 13).

[39] Latham & Watkins (n 7).

[40] Clayton Utz (n 14).

[41] Getty Images (n 1) .

[42] Pinsent Masons, ‘Getty Images v Stability AI: why the remaining copyright claims matter’ (Pinsent Masons, 4 December 2025) https://www.pinsentmasons.com/out-law/analysis/getty-images-v-stability-ai-copyright-claims-significance accessed 10 February 2026.

[43] Associated Press (n 26).

[44] ABC News (n 29).

[45] Reuters, ‘Getty Images largely loses landmark UK lawsuit over AI image generator’ (Reuters, 4 November 2025) https://www.reuters.com/sustainability/boards-policy-regulation/getty-images-largely-loses-landmark-uk-lawsuit-over-ai-image-generator-2025-11-04/ accessed 10 February 2026.

[46] Mishcon de Reya (n 3).

[47] Osborne Clarke (n 34).

[48] William Fry (n 36).

[49] Reuters (n 45).

[50] Clayton Utz (n 14).

[51] Latham & Watkins (n 7).

[52] Osborne Clarke (n 34).

[53] Pinsent Masons (n 42).

[54] William Fry (n 36).

[55] ibid.

[56] Keystone Law (n 31).

[57] Clayton Utz (n 14).

[58]  Latham & Watkins (n 7).

[59] Latham & Watkins (n 7).

[60] Getty Images (n 1) -.

[61] ABC News (n 29).

[62] William Fry (n 36).

Related Posts

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top