UK High Court: landmark decision on copyright and Gen AI

Focus on secondary copyright infringement       

The judgment does not determine whether there was any copyright infringement involved in the training of the Stable Diffusion model itself because those allegations (of primary infringement) were withdrawn by Getty at the conclusion of the trial, due to insufficient evidence of relevant activity in the UK.  Claims as to infringement arising from outputs were also abandoned by Getty after the prompts alleged to have been used to generate examples of infringing output were blocked by Stability AI. Since the database rights claim was inherently linked to the primary infringement and outputs claims, this was not advanced by the time of trial. The key copyright question that remained for the judge was therefore whether Stability AI was liable for secondary copyright infringement, for importing, possessing or dealing with an article which it knew or had reason to believe was an ‘infringing copy’.

Intangibles

In order to determine the secondary infringement issue the court first held that electronic copies stored in an intangible medium (such as in the cloud or, as in this case, an AI model) are capable of being an ‘article’ for the purposes of acts amounting to secondary infringement.

The finding that infringing articles can be intangible is not surprising but still a welcome confirmation for rightsholders of what many have predicted since the same judge allowed Getty’s case on the issue to go forward to trial following summary judgment application by Stability AI in late 2023.  Had the outcome on this issue been different and, as the judge said, if it only applied to tangible articles it “would deprive authors of protection in circumstances where the copy is itself electronic and is then dealt with electronically”. 

Importation

Given that finding, it is also unsurprising that the court held that downloads of some versions of the Stable Diffusion model amounted to importation into the UK. 

This is a clarification which will be more significant in use cases which do not involve trained AI models, such as where copies of works are made in other jurisdictions and accessed by users in the UK from ‘the cloud’. 

Infringing copy

For the purposes of secondary infringement, an article must be an infringing copy (with the emphasis on ‘copy’), and the judge said that an article cannot be an infringing copy if it has never consisted of, stored or contained a copy of a copyright work. Since the final iteration of Stability AI’s model was trained on copyright works (outside the UK) but never stored or contained a copy, it could not be an infringing copy.

In finding that the Stable Diffusion model is not an infringing copy the judge rejected Getty’s argument that, in effect, the model was tainted by the fact that its development involved the reproduction of Getty copyright works.  Storing the images of those works and then exposing the model weights to them in the training process was not enough because, as the parties had agreed, the model weights for each version of Stable Diffusion never contained or stored an infringing copy.

Whilst the judge did not precisely rule on the point, the implication is that, if an AI model has ever contained a copy of a work, even transiently, the model could be an ‘infringing copy’ for the purposes of secondary infringement.   It is worth bearing in mind that not all generative AI models are developed or work in precisely the same way as Stable Diffusion, so it is possible that, on a different fact pattern, the decision may be different. Also, different parties in future litigation may not be prepared to concede that the model weights or other aspect of the model(s) then in issue do not amount to or contain copies, and then again, perhaps the result on the secondary infringement issue may be different.  Nevertheless, this decision will make it much harder for such arguments to succeed. 

Trade Mark infringement

The High Court held that the appearance of watermarks in images generated by early versions Stability AI’s ‘Stable Diffusion’ synthetic image generator infringed Getty’s “iStock” and “Getty Images” registered trade marks. The output images bearing the relevant trade mark (or a closely similar sign) would be perceived by the average consumer as the commercial communications of Stability AI because they were generated by its model over which it has control, despite arguments by Stability AI that it is users of Stable Diffusion who are responsible for conditioning the circumstances in which outputs are generated.

Comment

The result on the key issue of secondary copyright infringement is good news for developers whose AI models are produced and operate in essentially the same way as Stable Diffusion.  They now have confirmation that, provided no training activities are taking place in the UK, they can commercialise their models in the UK.     

UK rightsholders, on the other hand, are now in a worse position in relation to AI models trained outside the UK than EU rightsholders, who have the benefit of the obligations on developers in the EU AI Act, to demonstrate compliance with EU copyright law, even where the model is trained outside the EU.

Whilst this decision highlights the fact that UK rightsholders have no mechanism to prevent models trained outside the UK from being commercialised in the UK, as mentioned at the start, the judgment did not rule on whether training the model amounts to copyright infringement. UK rightsholders are therefore still in a better position than EU counterparts to prevent use of their works for AI training in the UK without their express consent, at least while the UK exception for text and data mining remains limited to non-commercial use, pending any changes to UK legislation. 

In the wider (political) context, this judgment is likely to give fresh impetus to lobbying efforts of both developers and rightsholders, in particular on the key proposals set out in the UK Government consultation on AI and Copyright, such as the Government’s proposed introduction of a commercial text and data mining exception, and the Government's request for input on whether any levers should be applied, to encourage developers to comply with UK copyright law where the training activity is outside the UK (which would bring the UK more in line with the EU AI Act). For more on the UK copyright consultation see our earlier article here.

Authored by Alastair Shaw and Penny Thornton.