As quickly as that, the Getty Images v Stability AI trial has come and gone. Having watched the trial unfold, here’s my commentary on what happened and where it leaves UK law.
Getty brought a claim in the England and Wales High Court against Stability AI for primary copyright, secondary copyright, database, and trademark infringement. Getty claimed that Stable Diffusion was trained on vast quantities of Getty-licensed images without permission and could reproduce those images on request. Sometimes the images produced also included the Getty watermark, although both parties seemed to accept that this was unintentional.
This trial was full of unexpected turns – in its first week, it was almost derailed twice because of a niche point around the wording of the pleadings. That was resolved with an incredibly fast decision by the England and Wales Court of Appeal.
The decision confirmed that in this context, the wording of Getty’s pleaded trademark tarnishment case (which referenced ‘pornography’) was not broad enough to encompass child sexual abuse material. The appeal court’s decision gave the legal world some nice practice points around the importance of being specific about your pleaded case at an early stage, particularly if the other side has explicitly requested further information.
The most important unexpected development was Getty dropping its primary copyright infringement and database infringement claims on the first morning of closing submissions.
As a result, this case will no longer give rise to a decision on whether the Stable Diffusion model infringed Getty’s copyright either by;
1) Training on Getty images in the UK; or
2) Substantially reproducing Getty images in the UK.
This late-stage amendment was likely prompted by difficulties with jurisdiction – although there has clearly been reproduction of copyright-protected works in the process of training Stable Diffusion, most of that training took place in the US (where a parallel claim is ongoing).
Getty struggled throughout the UK proceedings to establish that any copyright-restricted actions took place within the jurisdiction.
As a result, by the end of the trial, Getty’s claims were slimmed down to trademark infringement (identical use, confusingly similar use, and tarnishment) and secondary copyright infringement (which covers downstream dealings with copyright-protected works, such as importation). There was also a passing off claim, which largely stands and falls with the trademark claims.
Looking at both in more detail:
Trademarks: Before proceedings began, I was persuaded that Getty would establish the necessary detriment to support its trademark infringement claims. However, Stability built a convincing argument that, as people better understand the capabilities and typical outputs of AI models, it is less and less likely that the occurrence of a watermark will give consumers a trade origin message.
There was a moment where the trademark case could have been saved by the UK Supreme Court’s ruling in Iconix v Dreampairs concerning post-sale confusion (which was handed down mid-trial).
Stability swiftly frustrated that point by showing that post-sale use is unlikely to ever transpire in a typical AI use-case, where watermarked outputs are unintentional byproducts and disregarded.
Given that the “mischief” remedied by trademark infringement is tied to consumer perception, the trademark case is also disproportionately affected by the “contrivance” issue, which Stability reiterated throughout. In short, the images selected by Getty to evidence their case do not represent real life. They are artificially generated to support the litigation, using prompts selected by Getty’s lawyers. This undermines Getty’s argument that the distinctiveness and reputation of their trademarks are being damaged by the operation of the Stable Diffusion model in real life.
All of this ends up being very tricky for Getty, especially where its primary copyright infringement case has failed on the evidence.
There is clearly a harm to remedy (in its defence, Stability admits that at least some copyright-protected works were used in training version 2), but it does not appear that it is the type of harm which trademark infringement is designed to remedy.
Copyright: That takes us neatly to the copyright infringement case. There was a lot of discussion from both sides of a classic Lord Bingham quote, which I think gives a helpful insight into Getty’s copyright complaints. It explains why Getty’s secondary copyright infringement claim remains so important, even though the primary claim has fallen away.
In the famous Designers Guild v Russell Williams case, Lord Bingham confirmed that the owner of an original work shall enjoy for a limited period “the exclusive right to copy that work”, and for that period “no one else may for a season reap what the copyright owner has sown”. Clearly, Getty object to Stability’s “reaping” the benefit of its copyright-protected works for model training, whether this was conducted in the UK (primary infringement) or outside the UK and later imported or sold into the UK (secondary infringement).
Stability tried to explain this away by focusing on the fact that this quote relates to copies, and copies are not retained within the Stable Diffusion model.
With respect, I think this misses the wider principle Lord Bingham is articulating – there is value in original works and copyright ensures that, for a period, the author is the only person entitled to exploit that value.
In this case, Stability has exploited copyright-protected works for their huge value in training model weights. In defending the secondary copyright claim, Stability brought into issue the composition of the end-product, emphasising that it does not retain any copied works.
In my view, for UK copyright law to maintain its integrity, this should be irrelevant – extracting value from protected works is an act reserved to the copyright owner, whether those works are retained or discarded. Doing it requires permission, regardless of whether the act takes place in the UK (primary infringement) or the resulting benefit is sold to UK consumers (secondary infringement).
Whether the judge [Mrs Justice Joanna Smith] agrees with this approach remains to be seen – what is clear however is that her decision on secondary copyright infringement will give an important insight into whether the courts are willing to create a loophole in UK copyright law (train abroad and you avoid liability) to service the government’s ambition to remain competitive on the global AI stage.
Any decision by the High Court is, of course, subject to appeal, and with so much funding behind these parties, we can expect the saga to continue long beyond this first instance decision, which is due in October.
Don’t, however, underestimate the importance of this moment – although (given its appealable nature) the decision might not change the approach taken by businesses in the short term, it will set the stage for the AI regulation we are expecting early next year.
It will force the government to get off the fence and decide whether its priority is protecting creators or fostering AI development.
Other players have already made their move – the EU AI Act has confirmed the extraterritorial effect of the EU copyright opt-out regime (largely understood to operate in favour of creators), whilst the US fair use rules have recently been confirmed flexible enough to legitimise AI training on copyright works (developer-friendly). It is time for the UK to make its call.
Are we the home of tech or creatives? We can no longer be both.