Enforcing IP against AI training across borders: UK guidance, Australian context

Where do rights owners stand when their material is used to train generative AI? 13 min read

The Australian Government recently announced it will not introduce a 'text and data mining exception' into Australia's copyright regime. This means that, in most circumstances, reproductions and extractions of copyright-protected material done without permission may amount to copyright infringement—strengthening copyright owners' claim to compensation if their copyright-protected material is used to train generative artificial intelligence (AI) models in Australia.

The potential for generative AI models to infringe the rights of intellectual property (IP) owners of material used as training data is a hot topic for litigation overseas. Among several recent decisions, the decision of the High Court of England and Wales in Getty Images v Stability AI is likely to be particularly instructive. Not only is this one of the first full decisions internationally on whether an AI developer has infringed content owners' IP rights by training an AI model on creative content without consent, the reasoning is particularly relevant for Australian innovators and content owners because of similarities to UK copyright law as it applies to AI.

In this Insight, we unpack the implications of both developments for owners of IP rights and developers of AI models in Australia.  

Who in your organisation needs to know about this?

• IP portfolio managers
• Software R&D managers
• AI model developers
• Legal counsel

Key takeaways 

• Australia has ruled out a text and data mining exception to copyright law, to ensure creators are fairly compensated if their copyright-protected material is used for training AI models.
• The UK Getty Images v Stability AI decision illustrates the complexities that arise from jurisdictional differences in legal rules and the territorial nature of IP law for copyright owners looking to enforce their rights.
• In Getty Images, the copyright owner, Getty Images, might have had a good case on reproduction of copyright works in the training data for generative AI because the UK (like Australia) does not have a general 'fair use' doctrine, and does not have a (commercial) text and data mining exception. However, it faced difficulty identifying the jurisdictions in which allegedly infringing acts had taken place.
• The court in Getty Images also did not accept that a generative AI model (once trained) stored copies of the works to which the model had been exposed in the training data. The result illustrates the potentially narrow protection for rights owners, even in a jurisdiction perceived as 'rightsholder friendly'.

TDM exception: what is it and why don't we have one?

A 'text and data mining' (TDM) exception is part of the copyright law of several major jurisdictions overseas, including the EU, Japan and the UK. In the US, the 'fair use' doctrine can apply to use of copyright material in the course of TDM.

The application and operation of the TDM exception varies by jurisdiction. For example, as implemented in the EU's Directive on Copyright in the Digital Single Market (Directive (EU) 2019/790), the TDM exception permits reproductions and extractions of lawfully accessible copyright-protected works and other subject matter for the purpose of 'any automated analytical technique aimed at analysing text and data in digital form in order to generate information which includes but is not limited to patterns, trends and correlations'. However, where TDM is undertaken for purposes other than scientific research, the copyright owner can exclude the TDM exception by an express reservation of their rights. By contrast, the UK has implemented a TDM exception on similar terms, but only for scientific research, not for commercial purposes.

In the era of generative AI, the TDM exception has become a talking point. An EU-style TDM exception (that includes commercial purposes) means that, if a copyright holder makes copyright-protected material lawfully accessible (eg by publishing it on the internet), they cannot charge an additional royalty for TDM activities. A prime example of a TDM activity is the collation of a training data set that can then be used to train or finetune a generative AI model.

In Australia, a TDM exception was proposed by the Productivity Commission in its interim report on harnessing data and digital technology, released on 5 August 2025. However, on 26 October 2025, the federal Attorney-General announced that the Australian Government 'will not be entertaining' a TDM exception. This announcement was made to provide certainty to Australian creators and to ensure they are fairly compensated if they make their copyright-protected material available for training AI models.

Australia is not alone in making this assessment: in the UK, the government considered expanding its TDM exception to commercial use but declined to pursue this change in 2023 after consultation.

Getty Images v Stability AI: a much-qualified win for rights owners

While the Australian Government's announcement clarifies one aspect of Australian copyright law applicable to AI, much remains to be seen as to the outcome for parties who wish to litigate such issues in Australia. The potential for generative AI models to infringe the IP in material used as training data has been subject of high-profile litigation in many overseas jurisdictions. However, the inconsistencies in copyright law between jurisdictions means that overseas decisions need to be treated with caution as guidance to how the same issues might be decided in Australia.

Two major decisions were handed down earlier this year in the US involving the unauthorised use of copyright-protected materials in training data for generative AI models: claims of copyright infringement brought against Meta by a group of authors whose works had been used to train Meta's Llama large language model,¹ and a class action brought by book authors against AI developer Anthropic,² which Anthropic ultimately settled for USD 1.5 billion. In both of these cases, the US courts held that training AI models on copyright-protected work, even without permission, could fall within the scope of the US 'fair use' exception to copyright infringement. However, the guidance that can be drawn from these cases for Australian rightsholders and developers is limited, because Australia does not have a general 'fair use' doctrine.

Fortunately, the Getty Images v Stability AI decision by the High Court of England and Wales is not only one of the first final decisions internationally in a dispute between rights holders and AI model developers, it also occurred against a legal framework that (on relevant issues) is similar to Australia's. The UK, too, does not have a US-styled general 'fair use' exception to copyright protection (and, instead, has purpose-limited 'fair dealing' exceptions). Although the UK has a TDM exception, it is limited to scientific research and does not cover commercial purposes.

Indeed, commentators have seen these features of UK copyright law as making it an attractive jurisdiction for IP rights owners to enforce their rights. However, the outcome of the case illustrates the qualifications to this view, which will be relevant to parties considering potential litigation in Australia.

Training claim: the importance of jurisdiction

The main plaintiff in the case was Getty Images, which owns or holds licences to the copyright of a large collection of images accessible to the public on its websites in a low-resolution format with a watermark, together with metadata relating to the image and keywords describing its subject matter. A customer may purchase a higher-resolution copy without a watermark.

The main defendant, Stability AI, developed Stable Diffusion, a generative AI image-generation tool.

Much of the public attention on the litigation arose from a series of comparisons showing that some images Stable Diffusion generated bore a remarkably close resemblance to Getty Images' content, and/or reproduced text that closely resembled the Getty Images watermark.

Stable Diffusion is a 'diffusion' model—it transforms an input prompt or seed image through probabilistic modelling that is iteratively sampled into a synthetic image. The basic architecture generates a random or 'stochastic' model, which is 'trained' through repeated exposure to human-made digital images, which iteratively refines the weights in the model. Once trained, users who implement the model only need to download the model weights—the training data is not needed for the trained model to operate and is usually not provided to users. To train a high-quality image generation model, it is necessary to use a structured set of training data that includes both images and high-quality data describing the image. Creating the data attaching to images requires significant human effort, so an existing set of images with descriptive data, such as that of Getty Images, would be an attractive data source

Stable Diffusion was trained on subsets of the LAION-5B dataset, a collection of URLs for images retrieved by a public web crawler, paired with alt-text captions describing the images' contents. It was common ground that this included Getty Images' content. The LAION-5B dataset was primarily prepared by LAION, another not-for-profit organisation.

In these circumstances, the stars seemed aligned for Getty Images to argue that the training activity infringed its copyright—especially when the case was brought in the UK, a jurisdiction with no commercial TDM exception and no general doctrine of fair use. However, jurisdiction proved to be a major factual obstacle: copyright is territorial, and Getty Images needed to identify an act of infringement that took place within the UK. The development and use of the training data, including the web scraping, downloading of the datasets and training of the models on the datasets, all took place in Germany and the US, and on Amazon Web Services' cloud hosting and processing resources physically located across various countries outside the UK.

In the result, the judge, The Honourable Mrs Justice Smith, accepted that an electronic copy of a work subject to copyright that is stored in an 'intangible' medium, such as a cloud service, is capable of being an infringing copy. She also accepted that the images were downloaded and stored, and copies of the images in the training data were temporarily made in the video random access memory (RAM) of the graphics processing units (GPUs) performing the training. However, this copying was all carried out as part of the model training outside the UK.

Getty Images dropped this claim during the final hearing.

Secondary infringement: a model is not an infringing copy of training data

Getty Images maintained its allegations that, by offering the Stable Diffusion service to British users through making the model weights for particular versions available for download, Stability AI was importing unlawful copies of Getty Images' images into the United Kingdom in breach of section 22 or 23 of the Copyright, Designs and Patents Act 1988 (UK) (the CDPA).

In order for a finding to be made that Stable Diffusion was an infringing copy of Getty Images' images, Getty Images needed to establish that the making of the model itself would have constituted an infringement of copyright if the model was made in the UK, following the relevant definitions in s27(3) of the CDPA.

The judge accepted that 'a large part of [the model's] 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'.³ However, this is not the same as the model storing the training data. She found that the Stable Diffusion models were not infringing copies because 'in its final iteration Stable Diffusion does not store or reproduce any Copyright Works and nor has it ever done so'.⁴

The judge noted it is technically possible for AI models to 'memorise' and be able to directly reproduce training data if a model's training focuses too much on a particular set of training data. However, there was no evidence of the Stable Diffusion models having this behaviour or being subject to the sort of training that would lead to 'memorisation'.

Trade mark infringement: some limited success for Getty Images

Getty Images also, and more successfully, alleged that Stability AI had infringed its word marks for ISTOCK and GETTY IMAGES and the Getty Images stylised logo, by Stable Diffusion generating images that included synthetic watermark-like signs.

The judge held that the preliminary issues for trade mark infringement were met: the watermark-like signs were being used 'in the course of trade' and 'in relation to goods and services' (as part of the 'commercial communication' of Stability AI to its customers). The judge also held that the 'digital imaging services' and 'downloadable digital illustrations' for which Getty Images held registrations were identical goods and services to the provision of synthetic image outputs that Stable Diffusion produced.

The judge found the ISTOCK mark had been infringed, due to the use of 'identical marks for identical goods or services' under s10(1) of the Trade Marks Act 1994 (UK) because there was evidence of Stable Diffusion images containing 'iStock' with no alterations. However, she determined that there was no infringement of the GETTY IMAGES word and logo marks under this section because the signs that Stable Diffusion generated were not completely identical. The AI-generated signs had different features such as extra letters, as illustrated in the below extract of an image that Stable Diffusion generated:

Further, Stability AI had conceded that the signs Stable Diffusion reproduced were similar to Getty Images' word and figurative marks, and the judge concluded that the use of these signs was likely to confuse a significant proportion of the relevant public. Consumers using Stable Diffusion would think that a generated image bearing an ISTOCK or GETTY IMAGES mark had been supplied by Getty Images or trained on images under a licence from Getty Images, and it did not matter that the signs were only encountered in a post-sale context. The use of a similar sign for similar or identical goods and services, which created a likelihood of confusion, constituted trade mark infringement under s10(2) of the UK Trade Marks Act.

Another head of infringement under UK trade mark law is dilution—use of a similar or identical sign that took unfair advantage of, or was detrimental to, the distinctive character or the repute of a registered mark. The judge found there was no trade mark infringement on this ground—there was limited evidence of images containing the watermarks being produced by real-world consumers, only earlier models of the Stable Diffusion AI had generated the synthetic watermarks, and there was no evidence that users would attempt to use Stable Diffusion to circumvent the need to pay for Getty Images' content.

As an overall comment, the judge emphasised at several points that the average consumer using Stable Diffusion would not consider that the responsibility of generating images with watermarks or watermark-like signs would fall solely or predominantly fall on themselves, regardless of the terms and conditions of use, particularly where the sign appeared unprompted or seemingly randomly.

While Getty Images was at least partly successful on trade mark infringement, the judge took pains to note that the analysis was fact-sensitive and would not necessarily be applicable in the same way to 'different watermarks generated on different images in response to different prompts'.⁵

Stability AI's liability

The judge found that Stability AI was responsible for the outputs of most of the Stable Diffusion models because of the level of control it had over the development, training (including data selection) and release of, filtering and guardrails for, and consumer access to the AI models. Stability AI's legal responsibility for the release of these versions of Stable Diffusion was common ground between the parties.

However, she excluded Stability AI from being liable for outputs created by the earlier versions of Stable Diffusion that were released by publication of the model weights on the GitHub or Hugging Face pages of CompVis, a third-party collaborator of Stability AI, since this was not done under Stability AI's control. Stability AI providing hyperlinks to the platforms on its website was nothing more than providing information about where a prospective user could locate the AI model weights and source code.

Implications for Australia

In the same month as the UK decision in the Getty Images v Stability AI case, the Munich Regional Court in Germany handed down a judgment in the GEMA v OpenAI case,⁶ which also concerned 'memorisation' by a generative AI model of training material, but this time concerning song lyrics. In contrast to the limited success of Getty Images in the UK case, GEMA (the German music collecting society) succeeded in its copyright claims, and convinced the court that OpenAI's conduct was not protected by Germany's EU-style TDM exception.

These contrasting results illustrate the significant impact of jurisdictional differences on outcomes, which may be difficult to predict. So, how might the facts of Getty Images v Stability AI have played out if the case had been run in Australia?

Due to the dispersed nature of the training activities and the territorial nature of IP rights, Getty Images would have faced the same issue regarding jurisdiction for any copyright infringement claims based on training.

In relation to the copyright and trade mark claims concerning the model itself, the outcome is likely to have been similar had the questions been asked in an Australian court. Australian law has similar, although not identical, provisions to the UK copyright and trade mark provisions considered in the case. The equivalent provisions in Australian legislation concerning secondary copyright infringement, ss37 and 38 of the Copyright Act 1968 (Cth), would similarly only make the importing of an AI model an infringement of copyright of the underlying training data if 'the making of the [model] would, if the [model] had been made in Australia by the importer, have constituted an infringement of the copyright'. It is also likely that an Australian court would have reached the same outcome regarding trade mark infringement under s120(1) and (2) of the Trade Marks Act 1995 (Cth)). A key difference with UK law is that Australian trade mark law does not provide for infringement through dilution. This sort of conduct is likely to be addressed through the misleading or deceptive conduct provisions of the Australian Consumer Law.

However, the law in Australia may yet change. While the Government has ruled out a TDM exception, it has indicated that the Copyright and Artificial Intelligence Reference Group within the Attorney-General's Department is currently exploring three priority areas in copyright and AI policy:

• licensing arrangements, including considering whether there should be a new paid collective licensing framework for AI.
• how to make copyright law clearer in terms of how it applies to AI-generated material.
• how to make enforcement of existing rights more cost-effective and simpler—eg through a new small claims forum.

Actions you can take now

• Assess your liability exposure: if you are using or co-developing an AI tool, and especially if you are presenting the tool for public use, discuss with us your potential exposure to liability for infringement of third-party IP rights.
• For developers of AI models:
  • Double-check your training data and training processes: consider the sources of training data, and where it was prepared, as copyright laws differ across jurisdictions.
  • Reassess what guardrails and parameters will prevent replicating IP-protected material belonging to other businesses: Stability AI was able to limit later models of Stable Diffusion from generating images with watermark-like signs by applying filters and guardrails to prevent replication of Getty Images' trade marks (although such filters will not be completely foolproof).

If you would like more information about the issues raised in this Insight, please contact any of the people below.