November 2025
UK’s First Generative AI Judgment: Getty Images v. Stability AI — Copyright Claims Dismissed
I. Overview & Milestone Significance
On November 4, 2025, the High Court of Justice of England and Wales issued its judgment in the case where plaintiff Getty Images sued Stability AI [1] , alleging that its image generation tool, “Stable Diffusion,” infringed its copyright and trademark rights. The Court found partial trademark infringement but dismissed the copyright secondary infringement claims. Regarding the core question of whether training and development constitutes copyright infringement, the court noted that, due to a lack of evidence proving Stable Diffusion's training and development occurred within the UK, the plaintiff withdrew this portion of the claims.
This case represents the UK's first judgment concerning intellectual property disputes involving generative artificial intelligence (AI) models. Although the legality of training data was not within the scope of the final judgment, the court still conducted a substantive review of whether the weights and outputs of generative AI models constitute copyright infringement, establishing important precedential significance for subsequent cases.
In other words, this case not only marks the UK courts' formal response to intellectual property disputes arising from generative AI, but also represents a major milestone as the world's first case in the trial stage to explicitly address copyright infringement through arguments about AI model outputs and weights, following multiple US court rulings on "Fair Use."
US courts recently issued rulings in June 2025 in the Anthropic [2] and Meta [3] cases, finding that the use of works by generative AI models for training purposes has a high degree of "transformativeness" and qualifies as fair use. In contrast, this UK court judgment takes a different argumentative approach from US fair use analysis, using computational science data to demonstrate that "Stable Diffusion models do not constitute infringing copies" as its core reasoning.
II. Copyright Infringement Dismissed: Model Is Not an Infringing Copy
The plaintiff alleged that, when developing the Stable Diffusion model, the defendant used copyright-protected images from its website as training data without authorization, constituting copying under Sections 17 and following of the Copyright, Designs and Patents Act 1988 (CDPA).
The court noted that the training dataset LAION-5B, used by the defendant, was automatically crawled from the public web (Common Crawl) by the German non-profit organization LAION e.V. Although the plaintiff claimed the dataset contained links to numerous images from its website, the court did not make findings regarding their exact quantity or content.
During training, the defendant would temporarily download these images for so-called "materialisation" processing. However, based on expert testimony from both parties, the court found that the Stable Diffusion model does not store any concrete images, retaining only statistical associations and weight parameters. The enormous capacity difference between the weight files (approximately 3.44 GB) and the training dataset (approximately 220 TB) demonstrates that the model cannot possibly contain original pixel data.
Accordingly, the court determined that Stable Diffusion only learns statistical features and lacks the ability to reproduce specific images, and therefore does not constitute a "copy" under the CDPA. The model weights do not constitute infringing copies; Stability AI's conduct does not constitute secondary infringement, and therefore the plaintiff's copyright infringement claims fail.
III. Trademark Infringement Finding: Limited to Specific Versions
Beyond copyright, the plaintiff also alleged that some images generated by the defendant contained the plaintiff's trademark watermarks, potentially constituting infringement under the Trade Marks Act 1994 (TMA).
The court noted that watermark appearances are mostly incidental and primarily result from users inputting relevant terms in prompts, rather than being inherent output characteristics of the model. However, after examining different model versions and generated samples, the court found that in a few specific instances, trademark infringement did occur: some output images from Stable Diffusion v1.x versions contained the plaintiff's trademark watermarks, constituting infringement under TMA Section 10(1) (identical mark and identical goods/services) and TMA Section 10(2) (similar mark, similar goods/services, with likelihood of confusion);
1. Some output images from Stable Diffusion v2.1 versions contained the plaintiff's trademark watermarks, constituting infringement under TMA Section 10(2).
2. Some output images from Stable Diffusion v2.1 versions contained the plaintiff's trademark watermarks, constituting infringement under TMA Section 10(2).
Regarding the plaintiff's additional claim under TMA Section 10(3) for dilution or unfair advantage, this was dismissed as the plaintiff failed to prove any reputation dilution or detriment.
IV. Limitations of Judgment Scope: Training Data Legality Questions Remain Unresolved
Notably, this judgment did not substantively address the legality of AI model training data. The court determined that Stability AI's training activities occurred outside the UK, and because the plaintiff could not prove that the infringing acts occurred within the UK, they withdrew claims regarding "AI model training infringement" and "database rights infringement" midway through the trial.
Therefore, the core of this judgment is limited to the nature of the Stable Diffusion model and its output content, rather than whether the training phase itself is lawful. In other words, while the court found that the model is not an infringing copy, it did not answer the substantive question of "whether using copyright-protected images as AI training data constitutes infringement." The judge explicitly stated in the judgment that the cross-border nature and technical complexity of AI model training exceed the scope of current copyright law, and related issues should be further regulated by the legislature.
This case's outcome provides short-term legal certainty for generative AI developers, confirming that model weights themselves do not constitute infringement, but also highlights the reality gap in current systems regarding clear regulations on training data sources, fair use of copyright, and cross-border application. In the future, how to balance AI innovation with copyright protection will still depend on legislative and international coordination progress.
[1] Getty Images (US), Inc. & Ors v. Stability AI Ltd [2025] EWHC 2863 (Ch).
[2] Jane Tsai (2025, November 25). The boundaries of fair use in AI training: Insights from Bartz et al. v. Anthropic PBC in the U.S. https://www.leetsai.com/the-boundaries-of-fair-use-in-ai-training-insights-from-bartz-et-al-v-anthropic-pbc-in-the-u-s?lang=en-US
[3] Jane Tsai (2025, November 25). The Boundaries of Fair Use in AI Training: Insights from Kadrey et al. v. Meta Platforms, Inc. in the U.S. https://www.leetsai.com/the-boundaries-of-fair-use-in-ai-training-insights-from-kadrey-et-al-v-meta-platforms-inc-in-the-u-s?lang=en-US
On November 4, 2025, the High Court of Justice of England and Wales issued its judgment in the case where plaintiff Getty Images sued Stability AI [1] , alleging that its image generation tool, “Stable Diffusion,” infringed its copyright and trademark rights. The Court found partial trademark infringement but dismissed the copyright secondary infringement claims. Regarding the core question of whether training and development constitutes copyright infringement, the court noted that, due to a lack of evidence proving Stable Diffusion's training and development occurred within the UK, the plaintiff withdrew this portion of the claims.
This case represents the UK's first judgment concerning intellectual property disputes involving generative artificial intelligence (AI) models. Although the legality of training data was not within the scope of the final judgment, the court still conducted a substantive review of whether the weights and outputs of generative AI models constitute copyright infringement, establishing important precedential significance for subsequent cases.
In other words, this case not only marks the UK courts' formal response to intellectual property disputes arising from generative AI, but also represents a major milestone as the world's first case in the trial stage to explicitly address copyright infringement through arguments about AI model outputs and weights, following multiple US court rulings on "Fair Use."
US courts recently issued rulings in June 2025 in the Anthropic [2] and Meta [3] cases, finding that the use of works by generative AI models for training purposes has a high degree of "transformativeness" and qualifies as fair use. In contrast, this UK court judgment takes a different argumentative approach from US fair use analysis, using computational science data to demonstrate that "Stable Diffusion models do not constitute infringing copies" as its core reasoning.
II. Copyright Infringement Dismissed: Model Is Not an Infringing Copy
The plaintiff alleged that, when developing the Stable Diffusion model, the defendant used copyright-protected images from its website as training data without authorization, constituting copying under Sections 17 and following of the Copyright, Designs and Patents Act 1988 (CDPA).
The court noted that the training dataset LAION-5B, used by the defendant, was automatically crawled from the public web (Common Crawl) by the German non-profit organization LAION e.V. Although the plaintiff claimed the dataset contained links to numerous images from its website, the court did not make findings regarding their exact quantity or content.
During training, the defendant would temporarily download these images for so-called "materialisation" processing. However, based on expert testimony from both parties, the court found that the Stable Diffusion model does not store any concrete images, retaining only statistical associations and weight parameters. The enormous capacity difference between the weight files (approximately 3.44 GB) and the training dataset (approximately 220 TB) demonstrates that the model cannot possibly contain original pixel data.
Accordingly, the court determined that Stable Diffusion only learns statistical features and lacks the ability to reproduce specific images, and therefore does not constitute a "copy" under the CDPA. The model weights do not constitute infringing copies; Stability AI's conduct does not constitute secondary infringement, and therefore the plaintiff's copyright infringement claims fail.
III. Trademark Infringement Finding: Limited to Specific Versions
Beyond copyright, the plaintiff also alleged that some images generated by the defendant contained the plaintiff's trademark watermarks, potentially constituting infringement under the Trade Marks Act 1994 (TMA).
The court noted that watermark appearances are mostly incidental and primarily result from users inputting relevant terms in prompts, rather than being inherent output characteristics of the model. However, after examining different model versions and generated samples, the court found that in a few specific instances, trademark infringement did occur: some output images from Stable Diffusion v1.x versions contained the plaintiff's trademark watermarks, constituting infringement under TMA Section 10(1) (identical mark and identical goods/services) and TMA Section 10(2) (similar mark, similar goods/services, with likelihood of confusion);
1. Some output images from Stable Diffusion v2.1 versions contained the plaintiff's trademark watermarks, constituting infringement under TMA Section 10(2).
2. Some output images from Stable Diffusion v2.1 versions contained the plaintiff's trademark watermarks, constituting infringement under TMA Section 10(2).
Regarding the plaintiff's additional claim under TMA Section 10(3) for dilution or unfair advantage, this was dismissed as the plaintiff failed to prove any reputation dilution or detriment.
IV. Limitations of Judgment Scope: Training Data Legality Questions Remain Unresolved
Notably, this judgment did not substantively address the legality of AI model training data. The court determined that Stability AI's training activities occurred outside the UK, and because the plaintiff could not prove that the infringing acts occurred within the UK, they withdrew claims regarding "AI model training infringement" and "database rights infringement" midway through the trial.
Therefore, the core of this judgment is limited to the nature of the Stable Diffusion model and its output content, rather than whether the training phase itself is lawful. In other words, while the court found that the model is not an infringing copy, it did not answer the substantive question of "whether using copyright-protected images as AI training data constitutes infringement." The judge explicitly stated in the judgment that the cross-border nature and technical complexity of AI model training exceed the scope of current copyright law, and related issues should be further regulated by the legislature.
This case's outcome provides short-term legal certainty for generative AI developers, confirming that model weights themselves do not constitute infringement, but also highlights the reality gap in current systems regarding clear regulations on training data sources, fair use of copyright, and cross-border application. In the future, how to balance AI innovation with copyright protection will still depend on legislative and international coordination progress.
[1] Getty Images (US), Inc. & Ors v. Stability AI Ltd [2025] EWHC 2863 (Ch).
[2] Jane Tsai (2025, November 25). The boundaries of fair use in AI training: Insights from Bartz et al. v. Anthropic PBC in the U.S. https://www.leetsai.com/the-boundaries-of-fair-use-in-ai-training-insights-from-bartz-et-al-v-anthropic-pbc-in-the-u-s?lang=en-US
[3] Jane Tsai (2025, November 25). The Boundaries of Fair Use in AI Training: Insights from Kadrey et al. v. Meta Platforms, Inc. in the U.S. https://www.leetsai.com/the-boundaries-of-fair-use-in-ai-training-insights-from-kadrey-et-al-v-meta-platforms-inc-in-the-u-s?lang=en-US
