Copyright law protects intellectual creations such as books, songs, newspaper articles and movies, subject to certain exceptions. The datasets ingested to train artificial intelligence (AI) systems often include copyrighted material. AI developers argue that the use of this copyrighted material falls within certain of the legal exceptions to copyright protection. 

Whether or not that is the case has become one of the most contentious legal issues in the rapidly evolving AI landscape. It has already resulted in several high-profile legal cases involving copyright holders and major AI developers.

One of the first of these cases is currently being heard in the UK, with the trial in Getty Images' case against Stability AI¹ commencing on Monday.

The decisions of the Courts in both the EU and the UK will establish important precedents on whether training AI models using copyrighted works constitutes a breach of copyright law, potentially reshaping how AI developers approach data collection and AI Model training, and how creators protect their copyright.

What is copyright law?

Copyright law concerns the legal protection of the rights of human creators by granting them exclusive control over their original works. This legal protection, in effect, ensures that people maintain control over their creations and can be financially compensated for any future use of these copyrighted works.

In the EU, copyright protection generally lasts for 70 years after the death of the author and is automatic upon creation of an original work. No registration is required. For copyright protection, works must demonstrate originality, which is defined as being the author's own intellectual creation.² Literary, dramatic and musical works, visual arts and serial publications such as magazines, newspapers and journals are all generally deemed to be protected by copyright.

Copyright holders have a "reproduction right", whereby any reproduction, whether temporary or permanent, direct or indirect, requires authorisation from the rights holder.^3 4 For press publishers, and the use of newspaper articles to train AI systems has been a particular area of focus, this right includes the right to authorise or prohibit the online use of press publications by news aggregators and search engines.⁵ 

How does AI interact with copyright law?

One of the main issues that has emerged in relation to AI and copyright law is that Generative AI (GenAI) systems require massive datasets for training, and much of this data inevitably includes copyrighted works that developers often use without explicit permission from rights holders and indeed, often without the knowledge of the rights holders.

In the US, the GenAI framework primarily revolves around the application of existing copyright laws, particularly the fair use doctrine, to the use of copyrighted materials in AI training datasets.

In the EU and the UK, the focus is often on introduced exceptions to general copyright protection the scope of which differ to the US fair use doctrine. This contrast presents difficulties for global organisations deploying GenAI, as they must navigate multiple legal frameworks. It can also present difficulties for copyright holders, as they may be met with jurisdictional defences – for example, one of Stability AI's defences is that the training and development of its product, Stable Diffusion, took place outside of the UK⁶.

An exception that is frequently relied upon by AI developers in the EU and the UK for text and data mining (TDM)⁷ is that it is being done as research for non-commercial purposes.⁸ For example, the scientific TDM exception has been found to apply where the purpose of the use was aimed at gaining new knowledge and the resulting dataset was made available to the public for free.⁹

Given that TDM by AI developers is typically for commercial purposes, a more relied on exception is that use of the copyrighted work is permitted unless rights holders have expressly reserved their rights through an opt-out mechanism.¹⁰

This opt-out must be made in an appropriate and machine-readable manner for content made publicly available online. There is uncertainty as to what precisely constitutes machine-readable. The legislation lists "metadata" and "terms and conditions of a website or service" as examples,¹¹ and obiter dictum comments of a German Court¹² suggest that a reservation, even if written solely in natural language, will meet the "machine-readable" requirement for a valid opt-out under TDM provisions.

An exception that was unsuccessfully raised in a recent case before the Courts¹³ was for temporary acts of reproduction that are transient or incidental and technically required and have no independent economic significance.¹⁴ This defence was unsuccessful as the German Court found that the reproduction was not incidental and was actively controlled as it was a result of deliberate programming.

Another exception which relates primarily to the copyrighted material for press publishers is the use of individual words or very short extracts of a press publication.¹⁵ This exception is now being considered by the Courts of Justice of the European Union in a case involving Google Ireland.¹⁶ In this case, Google Ireland argues, amongst other defences being maintained, that longer extracts from newspaper articles do not constitute an instance of making content available to the public or reproduction as the extracts are not reaching a different public to the original publication.

There are also other exceptions or limitations that EU member states have a discretion to introduce such as for the purpose of benefitting people with a disability, public security, religious celebrations or of caricature, parody or pastiche.¹⁷ In the Getty Images case, Stability AI argues that the images that resemble Getty's content are pastiche which is an exception under UK copyright law, albeit one which has not yet been analysed in detail by the UK Courts.

Where to next?

The legal landscape for GenAI is at a pivotal stage, with an increasing number of legal proceedings being brought by copyright holders against AI developers.

These cases underscore the growing need for legal clarity regarding the interaction between copyright law and AI. Court findings in favour of copyright holders could mean that licences are required before AI developers can use protected content in the UK and EU.

As AI continues to evolve and challenge, it is inevitable that many more cases will arise. The outcome of these ongoing proceedings and further updates to the relevant legislation will be crucial in defining the balance between innovation and the protection of copyright, ultimately shaping the future of AI development.

1. Getty Images v. Stability AI Case IL-2023-000007.↩
2. In Infopaq International A/S v. Danske Dagblades Forening (C-5/08) at [2] interpreting the application Article 2(a) of Directive 2001/29/EC (Infosoc Directive). ↩
3. Article 2 of the InfoSoc Directive, transposed in Ireland through section 39 of the Copyright and Related Rights Act 2000 (Copyright Act).↩
4. Article 3 of the InfoSoc Directive, transposed in Ireland through section 40 of the Copyright Act.↩
5. Article 15 of the Digital Single Market Directive (2019/790) (CDSM), transposed in Ireland through regulation 13(1) of European Union (Copyright and Related Rights in the Digital Single Market) Regulations 2021.↩
6. Getty Images has also brought proceedings against Stability AI in the US.↩
7. This is the use of automated analytical techniques to analyse text and data in digital form to generate information like patterns, trends and correlations.↩
8. Article 3(1) of the CDSM (transposed in Ireland through section 57(1) of the Copyright Act) permits TDM by research organisations and cultural heritage institutions for scientific research purposes, without requiring authorisation from rights holders. Section 29A of the UK Copyright, Designs and Patents Act 1988 which contains an exception for TDM for research for non-commercial purposes for the sole purpose of research for a non-commercial purpose, where the copy for TDM is where possible accompanied by a sufficient acknowledgment.↩
9. Robert Kneschke v. LAION e.V., Hamburg District Court, Case No. 310 O 227/23. This case is currently under appeal.↩
10. Article 4 of the CDSM.↩
11. Recital 18 of the CDSM.↩
12. Ibid 9.↩
13. Ibid 9.↩
14. Article 5(1) of the InfoSoc Directive (transposed by section 244 of the Copyright Act).↩>
15. Article 15(1)(iv) of the CDSM, transposed in Ireland through regulation 13(1)(c) of European Union (Copyright and Related Rights in the Digital Single Market) Regulations 2021.↩
16. Like Company v Google Ireland C-250/25.↩
17. A lengthy list of possible exceptions is at Article 5 of the InfoSoc Directive.↩