On 4 November 2025, the High Court of Justice of England and Wales delivered judgment in the matter of Getty Images v Stability AI [2025] EWHC 2863. The case related to the deployment of Stability AI's generative AI (GenAI) model, Stable Diffusion in the United Kingdom. Stable Diffusion is a GenAI tool that is able to create an image with a user generated prompt. The proceedings involved allegations of copyright infringement, trade mark infringement and passing off. Although Getty Images had some success based on its trade mark infringement allegations in relation to certain model versions and specific circumstances involving watermarks, it was not able to make out its claim of copyright infringement. 

At the time of writing, despite substantial litigation brought by copyright owners against AI developers in the United States and United Kingdom, no claims have yet been brought in Australia. One reason is that the most popular AI models have been developed and trained overseas, in jurisdictions such as in the United States and China. 

The decision in Getty Images v Stability AI highlights how the geographic disconnect between the location of alleged infringing acts and the forum in which a claim is brought can significantly affect the scope of those claims. During the course of the UK proceedings, Getty Images was required to abandon its claims of direct copyright infringement due to the geographic disconnect. The judgment offers valuable insights into the training of AI models and the way parties and courts may approach copyright infringement claims in Australia where those models are developed overseas.

Territorial limitation of copyright laws

Like the UK legislation, the Australian Copyright Act 1968 (Cth) (Copyright Act) operates nationally. Sections 36 and 101 of the Copyright Act make clear that copyright is infringed by a person who does or authorises the doing in Australia of any act comprised in the copyright without the licence of the copyright owner—this is referred to as 'direct infringement'.

In addition to direct infringement, sections 37 and 102 of the Copyright Act create liability for 'secondary infringement'.

These provisions provide that copyright is infringed by a person who, without the licence of the copyright owner, imports an article into Australia for the purpose of:

• selling, letting for hire, or by way of trade offering or exposing for sale or hire, the article;
• distributing the article for the purpose of trade, or for any other purpose to an extent that will affect prejudicially the owner of the copyright; or
• by way of trade exhibiting the article in public,

if the importer knew, or ought reasonably to have known, that the making of the article would, if made in Australia by the importer, have constituted an infringement of copyright. Liability is secondary because the infringer did not themselves engage in copying or transmitting the copyright material in Australia but rather imported an infringing article.

In short, the secondary infringement provisions create a hypothetical scenario whereby the creation of the imported item is treated as if it occurred in Australia, even though it was created elsewhere.

How the AI model was trained

The UK High Court found that Stability AI's model, known as Stable Diffusion, is a latent diffusion model that transforms text prompts into synthesised images by modelling probability distributions based on training data.

During training, the model adjusts internal parameters known as model weights—numerical values that are optimised through a process called 'stochastic gradient descent' (an optimisation technique used in machine learning). These model weights encode statistical patterns learned from the training data and form the backbone of the model’s ability to generate new images. Crucially, the court accepted that while Stability AI had created training datasets by scraping the internet without the consent of copyright owners, including those with Getty Images watermarks, the model does not store training data itself. Rather, images were downloaded, converted into a latent representation to which noise is added and used as an input from which the model predicts what should be there instead of the 'noise'. Based on how it performs, the weights are adjusted. The model learns patterns through these weights, with the model weights being approximately five orders of magnitude smaller than the training dataset.

Once the model has been trained, it operates by a process referred to as "inference". As noted in the judgment:

Inference does not require the use of any training data and the model itself does not store training data. However, a large part of its functionality is indirectly controlled via the training data. In other words, the way in which the network makes use of its multiple layers is the result of the training process. 

UK users generated outputs either by downloading the model (including the model weights) to run locally, or by accessing a cloud-based version through Stability's DreamStudio platform or Developer Platform API.

Getty Images' copyright claim originally centred on allegations that Stable Diffusion was trained on a dataset which included millions of Getty-owned images. Getty Images argued that this training constituted copyright infringement, as did the generation of images by users of Stable Diffusion that allegedly reproduced elements of its images.

Training-based infringement

However, a critical issue in the UK proceedings was jurisdiction—specifically, whether the allegedly infringing acts occurred within the UK. Stability AI asserted that all model training took place on servers located outside the UK, primarily on Amazon Web Services infrastructure in the United States. In an earlier decision, the court found that Getty's pleadings regarding direct infringement relied on inference and were underdeveloped the absence of direct evidence linking the training process to the UK. Ultimately, Getty was unable to prove that any relevant training occurred in the UK, and this part of its claim was abandoned.

The position is likely to be similar under Australian law.

If Stability AI downloaded and stored Getty’s images on servers located in Australia to create a training dataset, this could constitute direct infringement. Importantly, unlike the United States, Australia does not have a broad fair use exception. While the Copyright Act contains a defined suite of fair dealing exceptions, none of the current exceptions would apply to training an AI tool per se. Whether Australia’s copyright laws are too restrictive is currently being debated, including before the Productivity Commission, but as things stand, AI developers in Australia should only use copyright material that is licensed or risk an infringement claim.

As the UK litigation demonstrates, the location of the training—and particularly where copies of the training datasets are made—is critical. If the training occurred entirely offshore, Australian courts may lack jurisdiction over the direct infringing acts involving the unauthorised reproducing, publishing or directly communicating to the public copyright material for training and fine-tuning of AI tools.

Output-based infringement

Australian courts would have jurisdiction over AI outputs generated in Australia that infringe copyright, at least where those outputs are saved, made available online, or electronically transmitted to users here.

However, the Stability AI litigation demonstrates that, notwithstanding the court's jurisdiction, proving that outputs are infringing can be difficult.
In Australia, the general approach to evaluating whether one work is a reproduction of another is to identify the work in which copyright subsists, then identify the part taken or copied in the alleged infringing work, and finally assess whether the part taken constitutes a substantial part of the original work.

The test for substantiality is qualitative rather than merely quantitative. The classic test involves a two-step inquiry: 

• first, whether there is sufficient objective similarity between the copyright work and the alleged infringing work; and 
• second, whether there is a causal connection, meaning the respondent’s work was derived from the copyright work and not created independently. 

The first issue depends on a comparison of the copyright work and the output. The second depends on whether, as part of the training, the developers scraped a copy of the work.

In the context of GenAI, where vast amounts of data are used for training and outputs vary, proving direct infringement in relation to outputs at scale will likely be a burdensome evidentiary task.

In the UK dispute, Getty Images was required to establish ownership of the copyright and whether the outputs it relied upon reproduced a substantial part of those images. This raised several issues:

• Even if an output is derived in part from a copyright work, it may be sufficiently different that it lacks objective similarity or the similarities do not meet the substantiality requirement.
• Outputs are not static as GenAI models usually generate different outputs to the same prompts at different times. While 'memorisation' issues can arise—where the output closely resembles a copyright work used in training—this phenomenon may not be stable, raising questions about the extent of infringement. The concept of 'memorisation'—where a model retains and reproduces specific training images—is contentious. The UK court found that while GenAI models can memorise training images through 'overfitting', this occurs when the model uses its weights to memorise individual training images rather than representing general patterns. The court noted that memorisation occurs through 'overfitting' when the model uses its weights to memorise individual training images rather than representing general patterns, and is an undesired feature in machine learning that engineers try to avoid.
• As raised by Stability AI in the UK litigation, outputs can be influenced by prompts. Stability AI argued that alleged infringements were the result of a wilful contrivance by Getty Images, using prompts that effectively created the infringements.

Getty Images withdrew its claims of direct infringement in relation to outputs during the proceedings. The court noted that Getty Images had achieved substantial relief regarding its output claims because Stability AI had blocked the prompts that Getty alleged had been used to generate infringing examples.

Courts may also be reluctant to impose liability on developers where the infringing use is initiated by end users, particularly if safeguards (such as prompt filtering) are in place. Prompt filtering is arguably akin to content moderation, which in the case of Redbubble Ltd v Hells Angels Motorcycle Corporation (Australia) Pty Ltd [2024] FCAFC 15 was a relevant factor in mitigating a finding that Redbubble had flagrantly infringed certain Hells Angel trade marks, and overturning an award of additional damages. 

Secondary Infringement

Getty Images maintained and pursued its claim for secondary infringement under sections 22 and 23 of the Copyright, Designs and Patents Act 1988 (UK) (CDPA). This novel claim hinged on whether the Stable Diffusion model itself constitutes an 'infringing copy', and whether Stability AI imported or distributed it in the UK with the requisite knowledge of its infringing nature. The UK High Court ultimately dismissed this claim.

Section 22 of the CDPA provides:

The copyright in a work is infringed by a person who, without the licence of the copyright owner, imports into the United Kingdom, otherwise than for his private and domestic use, an article which is, and which he knows or has reason to believe is, an infringing copy of the work.
(emphasis added)

Section 23 of the CPDA provides:

The copyright in a work is infringed by a person who, without the licence of the copyright owner—

(a) possesses in the course of a business, 

(b) sells or lets for hire, or offers or exposes for sale or hire, 

(c) in the course of a business exhibits in public or distributes, or 

(d) distributes otherwise than in the course of a business to such an extent as to affect prejudicially the owner of the copyright, 

an article which is, and which he knows or has reason to believe is, an infringing copy of the work.

(emphasis added)

"Infringing copy" is defined at section 27 of the CPDA as follows:

(2) An article is an infringing copy if its making constituted an infringement of the copyright in the work in question. 

(3) An article is also an infringing copy if—

(a) it has been or is proposed to be imported into the United 
Kingdom, and 

(b) its making in the United Kingdom would have constituted an infringement of the copyright in the work in question, or a breach of an exclusive licence agreement relating to that work.

Getty Images contended that the Stability Diffusion AI model was an 'article' and an infringing copy because its creation involved copyright infringement, namely, the reproduction of copyright works into the training data, even though the model weights themselves do not store or reproduce the training images.

While the UK court found that the AI model could be an 'article', it found that an AI model which does not store or reproduce any copyright works cannot be an infringing copy, even if its development involved the reproduction of copyright works during training. The court agreed with Stability AI's argument that the CPDA requires the article itself to include a reproduction of the copyright work(s) in material form. As Stable Diffusion does not retain a copy of the training data, the court determined that there had been no secondary infringement.

Notably, the secondary infringement provisions of the Australian Copyright Act differ from the CPDA. Section 37 provides:

(1) Subject to Division 3, the copyright in a literary, dramatic, musical or artistic work is infringed by a person who, without the licence of the owner of the copyright, imports an article into Australia for the purpose of:

(a) selling, letting for hire, or by way of trade offering or exposing for sale or hire, the article;

(b) distributing the article:

(i) for the purpose of trade; or

(ii) for any other purpose to an extent that will affect prejudicially the owner of the copyright; or

(c) by way of trade exhibiting the article in public;

if the importer knew, or ought reasonably to have known, that the making of the article would, if the article had been made in Australia by the importer, have constituted an infringement of the copyright.

(2) In relation to an accessory to an article that is or includes a copy of a work, being a copy that was made without the licence of the owner of the copyright in the work in the country in which the copy was made, subsection (1) has effect as if the words “the importer knew, or ought reasonably to have known, that” were omitted.

Section 37 does not expressly import the requirement that the article itself is an 'infringing copy', unlike the UK provision. As such it is possible that the Australian provision could be construed more broadly than the analogous UK provision. 

However, consistent with the UK position, one key issue under Australian law is whether the model is an 'article'. In contrast to the UK, the Copyright Act includes a definition of 'article'—at least in s 115—which expressly includes reproductions or copies 'in electronic form' such as computer files. Relatedly, the definition of 'material form' in the Copyright Act includes 'any form (whether visible or not) of storage of the work … or a substantial part of the work …, (whether or not [it] can be reproduced)'.

Although considering different provisions in the Copyright Act, in Universal Music Australia Pty Ltd v Cooper [2005] FCA 972, it was held that MP3 files could be copies of sound recordings, given, amongst other things, that sounds can be reproduced from them. In that same case, which was determined before the definition of 'article' referred to above was included in relation to the civil provisions of the Act, Tamberlin J at [92]–[93] held that:

There was some debate regarding whether the packet of electronic data which is activated by clicking on the hyperlink on the website to produce a download of the sound recording from the remote website can be described as an “article”. In plain and ordinary English usage, it could clearly not be so considered. The essential notation of an “article” in the Oxford English Dictionary is that of a material thing or tangible object.

... the ordinary English meaning should be applied to these terms and that the sound recording or bundle of electronic impulses in the form of an electronic hyperlink is not within even the most extensive definition of the expression “article”. In substance, what is being offered is a trigger mechanism for the activation of the free transmission of a packet of electronic data from the remote website to the user’s website. This does not fall within s 103. If an extended definition had been intended for the term “article”, as opposed to its ordinary English meaning, it would have been a simple matter for the draftsperson to draft the provision in such terms.

Consideration in Australia will also be focussed on whether the AI developer knew, or ought reasonably to have known, that the making of the article (i.e., the AI model) would, if made in Australia by them, have constituted an infringement of copyright. It remains a novel and open question under Australian law whether creating and using an infringing database to train an AI model would be considered to be a component of ‘making’ that model.

The secondary infringement provisions were not originally designed to deal with these types of scenarios but rather the importation of tangible things such as CDs. Whether courts are prepared to extend the application of the provisions to these new digital uses of copyright material or whether the Copyright Act needs to be updated to deal with these scenarios are open questions at least in Australia.

Other evidentiary and legal hurdles

Another issue that emerged from the Getty Images v Stability AI judgment is the difficulty in substantiating claims of copyright subsistence and ownership over a large number of copyright works. Some of the issues that arose included:

• Getty Images was unable to establish a clear chain of title to certain copyright works (including because individuals were employed by different entities within the corporate group, and there was no relevant assignment of copyright between the entities) and
• Getty Images did not have standing to sue in respect of some copyright works on the basis that it was neither the owner or exclusive licensee of the works.

While similar issues of fact and law are often in dispute in the context copyright litigation, a matter such as Getty Images v Stability AI serves to highlight the importance of ensuring that copyright owners and licensees maintain accurate records and ensure that licence agreements are appropriately drafted to preserve the option of copyright enforcement.

Implications of the UK decision

The Getty Images litigation underscores the difficulties for Australian copyright owners in stopping or seeking to redress the use of their works in training AI models overseas. Given the territorial nature of copyright law, the distinction between the location of user access and the location of model infrastructure will be highly relevant to, if not determinative of, the types of copyright claims that will be viable.

Copyright owners must be prepared to articulate clear claims grounded in territorial jurisdiction and evidentiary support. Vague or inferential claims will be attacked and are unlikely to succeed, particularly where the technical infrastructure of the AI tool is physically located outside Australia.

For AI developers, the case highlights the importance of data provenance, jurisdictional clarity, and technical transparency. Australian developers should:

• Avoid downloading or storing third-party content without proper licensing.
• Maintain records of training datasets and model architecture.
• Implement guardrails to prevent generation of content resembling protected works.
• Monitor user prompts and outputs for potential infringement risks.

The decision regarding Getty Images' UK copyright claim will be of great interest to copyright owners and AI developers. The procedural and evidentiary hurdles that arose in that case are instructive in terms of what claims could be pursued by copyright owners against AI developers who trained their models outside Australia. While the Copyright Act may still offer avenues for enforcement, any success will depend on the ability to prove substantial copying, causal connection, and convincing a court that provisions developed for tangible 'articles', such as the secondary infringement provisions, apply to AI models or the parts of them that are downloaded in Australia. While Getty Images achieved limited success on trade mark infringement claims in the UK, the failure of the copyright claim highlights the significant challenges facing Australian copyright owners seeking to enforce their rights against AI developers who train their models overseas. As generative AI continues to evolve, Australian courts and legislators are faced with a Copyright Act that appears increasingly out of date in the age of AI.

MinterEllison's market leading intellectual property team has deep expertise in copyright disputes. If you’d like to understand how this decision and Australia’s copyright laws affect your business, we’re here to assist.