Getty Images v Stability AI UK Judgment: The Scope of the Rulings on Generative AI, Copyright and Trade Marks

On November 4, 2025, Mrs Justice Joanna Smith DBE of the High Court of Justice, Business and Property Courts of England and Wales, Intellectual Property List (ChD), handed down judgment in the litigation between Getty Images and Stability AI (claim number IL-2023-000007, neutral citation [2025] EWHC 2863 (Ch)). The case concerned the development and provision of the image generation AI model Stable Diffusion, in which copyright infringement, database right infringement, trade mark infringement and passing off were at issue, and it is the first case in which an English court has ruled on the relationship between the training and provision of generative AI models and intellectual property rights. However, Getty Images withdrew major parts of its claims toward the end of the trial, and the judgment ruled not on the original claims as a whole but only on some of the issues that remained after they were narrowed. This article distinguishes the claims that were withdrawn from those actually decided, organizes the parts upheld and the parts dismissed, and considers the practical issues Japanese companies should check.

What Was at Issue (Course of the Litigation and the Withdrawn Claims)

Getty Images commenced this action on January 16, 2023. Its original claims included copyright infringement by the training and development activities themselves, on the ground that images from Getty Images and iStock had been included without permission in the training dataset for Stable Diffusion (part of LAION-5B) (Training and Development Claim); copyright infringement by outputs, on the ground that generated images reproduced Getty Images' works (Outputs Claim); and database right infringement ancillary to these (Database Rights Infringement Claim).

However, shortly before the close of the trial in June 2025, Getty Images withdrew all of these claims. The reasons were that there was no evidence that the training and development activities had taken place in the United Kingdom, that Stability had already blocked the prompts on which the output copyright infringement claim was based so that the purpose of the relief had been substantially achieved, and that the database right infringement claim was ancillary to these claims. Accordingly, this judgment does not rule on the issue of whether the act of training an AI model itself constitutes copyright infringement. I think this is the most important point to note in understanding the scope of the judgment.

What ultimately remained for trial were (i) trade mark infringement (sections 10(1), 10(2) and 10(3) of the Trade Marks Act 1994) and passing off where synthetic images containing the Getty Images or iStock trade marks are generated through ordinary use of Stable Diffusion, and (ii) secondary infringement (sections 22 and 23 of the CDPA), on the basis that the importation, possession and distribution of Stable Diffusion amounted to dealing in an "infringing copy" within the meaning of section 27(3) of the Copyright, Designs and Patents Act (CDPA). Getty Images did not allege that Stable Diffusion itself contained or stored copies of Getty Images' works; rather, it framed its case as the model itself being an infringing copy because the works were reproduced in the process of creating the model's weights.

The Court's Rulings on Trade Mark Infringement and Passing Off

The court first considered the scope of Stability's direct liability. The early versions of Stable Diffusion (v1.x) were developed through joint work between Stability and the academic research group CompVis and others, and were released on CompVis's GitHub and Hugging Face pages; while the court accepted that Stability had participated in and supported CompVis's development and release activities, it held that it could not go so far as to find that Stability was in a position to bear direct tortious liability for the release of v1.x.

Trade mark infringement was upheld only to a limited extent, after distinguishing between model versions. The claim under section 10(1) (identical sign and identical goods or services) was upheld only in respect of the fact that users accessing v1.x via DreamStudio or the API generated images containing iStock watermarks, and the claim under that subsection was dismissed with respect to Getty Images' own watermarks. The claim under section 10(2) (likelihood of confusion) was upheld in respect of the same v1.x iStock watermarks, and also in respect of the fact that users of v2.x generated images containing Getty Images watermarks. On the other hand, the claim under section 10(3), which addresses dilution of and free-riding on a brand, was dismissed in its entirety, and for SD XL and v1.6, trade mark infringement was not established for either version on the ground that there was no evidence at all that users had actually generated watermarked images. As for passing off, in light of the fact that neither party made additional submissions, the court took the approach of not ruling on it, as there was nothing to add to its rulings on trade mark infringement. In addition, on the issues of ownership of the copyright works and whether certain license agreements amounted to exclusive licenses, the outcome was split between upheld and dismissed depending on the subject.

Are the Model Weights an "Infringing Copy"?

The central issue in this judgment was whether secondary infringement was established, that is, whether the model weights of Stable Diffusion are an "infringing copy" within the meaning of section 27 of the CDPA. The court first held that an "article" under section 27 is not limited to tangible objects and may include intangible copies stored in electronic form. However, this holding did not determine the outcome of the claim. The court held that for an article to be an "infringing copy," there must have been a moment at which the article contained a copy of the work (following the analytical framework in Sony Computer Entertainment Inc v Ball), and found that, even though the model weights of Stable Diffusion changed as a result of exposure to Getty Images' works during training, the finished weights themselves have never stored or reproduced the works and will not do so in the future. As to Getty Images' argument that the time at which the training images were copied and the time at which the model was completed were simultaneous, the court rejected the understanding that "the model becomes an infringing copy at the moment it is made" as mistaken. As a result of this ruling, the secondary infringement claim was dismissed.

At the consequentials hearing in December 2025, Getty Images applied for permission to appeal the part of the judgment dismissing the secondary infringement claim, and the court granted permission. The reason given for granting permission was that the interpretation of "infringing copy" is an important, unprecedented question of law that may have wide implications for intangible articles such as generative AI models.

What the Judgment Did Not Decide

Some articles introducing this judgment carry headlines suggesting that the training of generative AI was held to be lawful or unlawful, but such an understanding cannot be said to be accurate. As matters the judgment did not decide, at least the following points need to be distinguished. Copyright infringement by the training and development activities themselves (Training and Development Claim) was not decided because Getty Images withdrew it. The claim that output images reproduced Getty Images' works (Outputs Claim) was likewise withdrawn and is outside the scope of the ruling. The database right infringement claim was also withdrawn. On passing off, the court reserved judgment. The court also stated expressly that it would make no finding on the number of works used for training.

Accordingly, what can be read from this judgment is that "in this case, there was no evidence that the training and development of Stable Diffusion took place in the United Kingdom" and that "the model weights themselves were held not to be an infringing copy under the CDPA, so long as they neither store nor reproduce the works"; no ruling was given on the lawfulness of the act of training generative AI in general.

Implications for Japanese Companies

This judgment concerns the interpretation of the UK CDPA and Trade Marks Act and cannot be read as a consequence under the Japanese Copyright Act. Japanese law has its own limitation on rights for the use of works for AI training, Article 30-4 of the Copyright Act, which adopts a structure permitting, under certain requirements, uses not aimed at enjoying the works. In contrast, this judgment did not presuppose any such express exception for training data, but determined whether the model weights are a copy of the works by interpreting existing provisions on reproduction, infringing copies and secondary infringement. Because the institutional frameworks differ, I think it is not appropriate to carry the conclusion of this judgment over as it is to situations in which Article 30-4 applies under Japanese law. Checkpoints for confirming the relationship between training data and copyright from the perspective of Japanese law are organized in Generative AI, Copyright and Training Data Checkpoints.

That said, there are several implications for corporate practice. Companies that use generative AI services to create marketing materials and image content are expected, as a practical response, to check what filtering the vendor performs at the output stage, in light of the possibility that other companies' trade marks or watermarks may appear in the generated content. In light of the fact that this judgment found trade mark infringement only for generation via DreamStudio or the API and made no similar finding for the mode of downloading and using the model, the perspective of separately checking the vendor's control systems for each mode of provision is useful. In contracts with AI vendors, it is practically important to check the sources of the model's training data, the status of rights clearance, and the allocation of responsibility for generated content. Perspectives for reviewing license agreements are explained in Intellectual Property and Scope of Use to Look at When Reviewing License Agreements, and items to check in intellectual property ownership clauses in Checkpoints for Reviewing Intellectual Property Ownership Clauses.

Furthermore, the ruling in this judgment that "the model weights are not an infringing copy" may be revisited as a result of Getty Images' appeal. In procurement review of AI vendors and in the design of contract clauses, it is premature to treat this ruling as settled law, and I think it is necessary to update one's assessment while monitoring the outcome of the appeal.

Developments to Watch

The information that could be confirmed as of the date of writing (July 12, 2026) is as follows. At the consequentials hearing held on December 16 and 17, 2025, Mrs Justice Joanna Smith granted Getty Images permission to appeal the part of the judgment dismissing the secondary infringement claim, and Getty Images was required to file its appellant's notice by February 3, 2026. On the other hand, the permission to appeal sought by Stability concerning the findings of trade mark infringement was refused, and for Stability to pursue an appeal, it would need to apply for permission itself to the Court of Appeal.

However, whether Getty Images actually filed its appellant's notice by the deadline, whether a hearing date has been set in the Court of Appeal, and whether Stability has applied for permission to appeal on its own could not be confirmed as of the date of writing. Accordingly, whether this judgment has become final has not been confirmed, and companies relying on it as a premise for procurement review or contract design need to keep monitoring the progress of the appeal.