Getty Images vs Stability AI: The Copyright Battle Set to Shape the Future of AI

This last week has seen closing submissions being made in the case of Getty Images (US) Inc vs Stability AI Ltd. It would not be an exaggeration to say that this is probably the most significant copyright case in the last decade.

Background to the case

At its heart is whether, as Getty claim, the scraping of millions of images from its website, without consent, and using them to train Stability’s AI model “Stable Diffusion”, and then the subsequent outputs from Stable Diffusion, infringe the copyright Getty own in those images.

Now that the trial is over, a judgment would normally be expected in three months, so by October 2025. The judge, Joanna Smith J, is developing a high-profile in so called soft IP cases (broadly those that cover for example trade marks and copyright rather than the more “sciencey” patent cases), having previously had her judgment in the Lidl v Tesco litigation where she found Tesco’s yellow and blue Clubcard branding infringed Lidl’s trade marks and passed them off upheld by the Court of Appeal. An appeal in this Getty case seems inevitable given what is at stake.

The question of how copyright laws are going to be applied in the development of AI has achieved global profile in recent years. Significant litigation (in addition to the Getty case in the UK) is underway in several other jurisdictions, including the USA. High profile UK artists such as Elton John and Paul McCartney have spoken publicly about their alarm over AI “ripping off” artists. On the other hand, there is political pressure for the UK to be at the forefront of AI development not least in its search for economic growth, and there have been suggestions from some politicians of reform of the law to enable this to happen, such suggestions often not being particularly well informed about how copyright law actually works.

The subject Getty case is almost several trials in one – there are substantive disputes involving a number of key IP rights, namely infringement of trade marks, passing off, infringement of copyright, infringement of copyright’s cousin database rights, and associated “safe harbour” defences. This is a very high-level summary, as many of the issues, particularly round the copyright and database rights claim, are complex.

What are the claims?

This article focuses on the copyright and database right claims and their associated defences, as those issues are the most pertinent to the relationship of our copyright laws to AI.

There are really three different but related copyright claims being advanced by Getty;

• Input
• Output
• Deployment

To establish the UK court’s jurisdiction, Getty needed to be able to show that there was downloading and storing of copyright works on hardware located in the UK as part of the process of training Stable Diffusion. Stability having done so in “the cloud” is unlikely to be enough. The problem for Getty here is that this is most likely to have taken place in the USA. But what if Stability had employees based in the UK doing this?

It would appear from the closing submissions that Getty may not have focussed on this aspect of their case by the end of trial and after the witnesses had been cross examined. We will have to wait for the judgement to find out, and obviously interested commentators such as myself are not privy to Getty’s thinking about its strategy and tactics in the case.

For claimants generally, although a case based on inputs is perhaps the most obvious way or first thought about how to advance a copyright AI case, the jurisdictional problems are likely to be significant where the AI models being trained are overseas, notably in the USA and China, and on cloud-based networks.

Output

This aspect of the case is all about whether the images then produced by Stable Diffusion infringe Getty’s copyright. Getty need to prove that a substantial part of their copyright works were reproduced in the output images created by Stable Diffusion (Stability argue that the images produced do not reflect “a substantial part of the intellectual creativity” of the photographer who took the original photograph). The actual legal test talks to “reproducing the work in any material form”, which has been interpreted by the courts as reproducing the whole or a substantial part”.

We will have to wait for the judgment to see if this subtle difference in language is intended to obscure what is actually an infringement, although it is understood from the closing submissions that this aspect of the case has also not been pressed by Getty.

Deployment

For most neutral observers, this is where it is probably felt that Getty have the best chance of success. The infringement case here is known as “secondary infringement”, which occurs where an infringer, (here, Stability), without licence of the copyright owner imports into the UK an article which is, and which he/she knows or has reason to believe is, an infringing copy of the copyright work.

The latter knowledge aspect of this test is likely to be easier to establish. It is the first aspect that still presents some serious challenges – does the word “article” cover something that is intangible such as an AI model? And can something be an “infringing copy” even if the article (assuming Stability is found to be an article) does not have any “copies” retained inside it?

Article

Stability’s argument is that the word “article” was never meant to mean anything other than a physical article one can touch hold and see. Getty on the other hand argues that legislation has always been read in light of subsequent technological developments. This sounds like a powerful argument – the law almost always lags behind technological developments – see for example how copyright law has adapted to the rise of the internet over the last two decades, with on line video platforms such as YouTube.

Infringing Copy

Getty say that infringing reproductions of copyright works were necessary to train Stable Diffusion, notwithstanding that such copies were transient in nature and not retained in the finished article. The relevant legislation (Copyright, Designs and Patents Act 1988) provides that copying can include the making of transient copies.

Legal defences to the claims

Stability rely on what are known as “E Commerce Safe Harbours”, which although deriving from an EU Directive take effect in the UK, even after Brexit. The particular defence likely to be most critical is the “hosting defence”. In outline, provided an intermediary operates within certain parameters outlined in the relevant legislation, it cannot be an infringer.

Stability says it is like Google – it merely processes data entered by third parties (i.e. users in Stability’s case, advertisers and users in Google’s case). Stability says it plays no active role in creating a new image and has no responsibility for its users’ actions. It is users they say who upload any infringing image. Thus, Stability seeks to shift responsibility onto its users, who one assumes might also be described as customers, saying it knows nothing of specific acts they may have committed.

Countering this defence, Getty says Stability is not an “intermediary” within the meaning of the hosting defence, as it is in a bi-partite not tri-partite relationship with its users, and in a bi-partite relationship one cannot be an intermediary. Further, Getty says, Stability goes well beyond a role which is merely technical, automatic and passive, which are characteristics required to fall within the safe harbour provided by the hosting defence.

Looking ahead to the judgment and its possible implications and consequences

Stability has said in its closing submissions at the end of the trial that if it loses, then UK users would not have access not only to Stable Diffusion, but also other internationally trained AI models – at least until the training of such models does not infringe a rights owners’ copyright works. Cue more hand waving from politicians agitating for a change in the law in search of economic growth.

But it would be equally true would it not that such an outcome would devalue the enormous value creative industries bring to the UK surely? According to a recent article in the Financial Times, the government has estimated that the UK’s creative industries generated £126bn in gross added value to the economy in 2022, roughly 5% of GDP, whilst employing around 2.4 million people. The FT question whether the “value add” of the AI industry will ever be of comparable scale.

Getty on the other hand says that a narrow interpretation of the relevant statute (i.e. a finding of non-infringement) would be problematic, and it is difficult not to see the force in that submission.

Some of those in favour of permitting scraping to further the advance of AI in the UK point to the “fair use” provisions in the Copyright, Designs and Patent Act 1988. Stability had originally relied on the “pastiche” fair dealing defence, arguing that the outputs generated are imitating the original work and are not a substitute nor does the imitation cause any harm to the original work, Generally, speaking the so called “fair use” provisions are for the purposes of “criticism review and reporting” before the introduction of the “caricature, parody, or pastiche” defence following the Hargreaves Review in 2012.

Whilst the trial presented an opportunity for judicial comment on the use and breadth of the pastiche defence, it is unsurprising that Stability chose to drop its pastiche defence during trial. The use of millions of images without payment for commercial enterprise does not fit easily within those provisions, as perhaps Stability and their lawyers recognised. As is always the case, it is for parliament to change the law, if a government feels change is needed, and for the courts to interpret the statutes in place at any one time.

It’s unlikely that this will be the last hearing in this case, and a trip to the Court of Appeal, and possibly even the Supreme Court, seems likely given the public interest