Copyright and Artificial Intelligence: does prohibiting text and data mining protect the UK creative industries?

As covered in detail in this article, one of the options raised in the original AI and Copyright Consultation was a broad text and data mining (TDM) exception, allowing AI companies to train models on copyright works that are scraped from the internet and other sources. Unsurprisingly, this received strong pushback from the creative industries, but there is an argument that a continued prohibition of commercial TDM in the UK will provide limited protection, as illustrated in the case of Getty Images v Stability AI. 

Before we discuss the case, it is useful to provide legal context. 

The legal geography of copyright 

Copyright protection is jurisdictional in nature. A UK rights holder does not hold global copyright. They hold UK copyright, and separately, subject to exceptions, whatever rights the laws of other countries choose to confer. 

The Berne Convention provides the international framework that provides baseline copyright protection for all signatory member states. However, these baseline rights focus on the reproduction of copyright works (the outputs of AI systems), rather than TDM (the inputs into AI systems). Even if this weren’t the case, there are several countries that are not members of the Berne Convention, so the rights of a copyright holder to exert international control over the use of their works are fragmented and non-uniform. 

This has direct consequences for AI training conducted abroad. Where the act of training (the scraping, processing, and ingestion of data that produces a working model) occurs outside of the UK, that activity will be governed by the jurisdiction in which it occurs and not UK copyright laws. 

Getty Images v Stability AI 

Getty Images issued UK proceedings in January 2023 against Stability AI over the development and use of Stability’s Stable Diffusion image-generation model. The claim was broad: Getty alleged that Stability had scraped millions of its copyright images without permission to train Stable Diffusion, and that the resulting model could generate outputs replicating Getty's copyright works. 

For our purposes, the important part of the claim is that Getty alleged direct (primary) infringement of their copyright works by Stability training their AI model on Getty’s images, in breach of the prohibition on commercial text and data mining. At trial, Getty conceded that the training and development of Stable Diffusion did not occur within the UK. As the training occurred outside the UK, the laws of the jurisdiction where training occurred applied, and so the UK courts did not have jurisdiction to rule on whether this constituted primary infringement.  

In essence, the UK’s commercial TDM restriction can be bypassed by training models in more favourable overseas jurisdictions. Preventing this would require the UK to enact extra-territorial legislation prohibiting the market entry of any model that would have infringed UK copyright had it been trained domestically. 

The practical limits of training AI abroad 

The more important lesson from Getty may be what it reveals about the limits of Stability AI’s strategy rather than its viability. Conducting training outside the UK reduces exposure to primary copyright claims but it does not eliminate legal risk entirely. A model trained in a permissive jurisdiction and then made available to UK users re-engages English law at the point of deployment, but now in relation to the outputs of the model.  

Provenance is becoming a commercial issue in its own right. As discussed in our article on international approaches to governing copyright and AI, the EU’s AI Act obligations require general-purpose AI providers to publish training data summaries for models deployed in the EU market, and accompany this with the ability for rights holders to opt-out from their works being used by AI systems. As a result, a model trained on unlicensed data in a permissive jurisdiction can face scrutiny when deployed in the EU. 

Additionally, the legal risk that jurisdiction shopping is designed to avoid can re-emerge as contractual and reputational risk at the point of commercial deployment when enterprise customers require IP indemnities and training data warranties. 

Therefore, it is arguable that the main effect of no TDM exception is to drive model training and generation abroad, cutting across the Government’s innovation strategy.  

---