Getty Images v Stability AI: the copyright ruling

Getty has lost on its claim of copyright infringement against Stability. Here, we unpack the ruling and what it might tell us about other possible cases. For the background to the claim, see here.

Getty's claim

Getty's allegation of secondary infringement is that Stability has imported into and distributed in the UK an 'article' (namely, the model weights) which is – and which Stability knows or has reason to believe is - an 'infringing copy' of Getty's copyright works contrary to sections 22 and 23 of the Copyright, Designs and Patents Act 1988 (CDPA). 

Getty's argument was firmly predicated on the fact that section 27 of the CDPA says that an article is an 'infringing copy' if its making would have constituted an infringement had it occurred in the UK. Getty did not argue that the Stable Diffusion model contained infringing copies.

The legal issues to be decided

There were two issues to be decided:

• Can an article be an 'infringing copy' if it does not contain a copy of the copyright works?
• Can an article be something intangible? 

The ruling

The judge ruled that whilst an 'article' may constitute an intangible object, an AI model such as Stable Diffusion, which does not and never has stored the Copyright Works, cannot be an 'infringing copy'. Getty's claim therefore failed. Below we consider each of the legal issues in turn. 

Can an article be an 'infringing copy' if it does not contain a copy of the copyright works?

The judge held that an 'article' cannot be an 'infringing copy' if it does not and never has stored the copyright works. To understand this part of the ruling fully, it is important to appreciate what Getty did and did not argue.

What Getty did not argue

Getty did not argue that the model weights store the copyright works. The model weights don’t store the copyright works directly or as pixel values, as can be seen if the size of the model weights is compared to the size of the training dataset. This was agreed by the parties' experts. 

However, it would have been open to Getty to argue that the model weights 'effectively' store the copyright works eg as patterns and statistics -arguments such as that the weights contain a 'residue' or some form of representation of the works or at least instructions on to how to render them. As one of the experts said, "…diffusion models learn the statistics of patterns which are associated with certain concepts found in the text labels applied to their training data ie.… they learn a probability distribution associated with certain concepts." 

We don't know why Getty did not run this argument, but there are some possible explanations:

• It might be because Getty could not point to a clearly memorised (nearly identical) output or any over-fitting that would suggest memorisation (over-fitting being where the network uses its weights or part of its weights to memorise individual training images, usually because of over-exposure to those images, rather than representing a large set of training images jointly). Getty did not assert that any work had been prioritised during the model's training such that over-fitting and memorisation would result - and (as the judge said) there was no evidence of any memorisation in fact (which is partly why Getty's claim of output infringement was dropped). While the judge did consider that overfitting and memorisation could occur, she did not draw any particular conclusions about what the implications of this would be. 
• It might be because of the difficulty of arguing that copyright in artistic works is infringed by the storing of 'patterns', 'statistics' and 'probability distributions'. The judge referred to the creation of the model weights as the "creation of information" and said that the article was an "intangible record of that information". In other words, she saw the model weights as more like a description of, or information about, an image than the image itself. It seems uncontroversial that a literary description of an artistic work, no matter how detailed, cannot infringe the copyright in that artistic work. The defendant's use has to represent the claimant's work in 'some real sense' (in the words of one of the leading textbooks, and adopted by the court in an earlier case). 

In future, it is possible that other copyright owners will seek to argue that the model weights 'effectively' store the copyright works - and that argument might gain some traction if the works alleged to be infringed are especially creative/unique, there is evidence of over-fitting and there is a clearer link between the input and output images. Model developers will no doubt already be factoring in the additional risk over-fitting presents to infringement and trying to stop it occurring. 

It is possible that another court (if asked the question) could find that model weights are more analogous to pixels or a form of compressed file format rather than information/descriptions such that they are deemed to contain copies of the works alleged to be infringed ie because the weights are a representation of a work in a real sense. Such a finding would likely depend quite heavily on the expert evidence in the subsequent case. However, if a court did make this finding, then it would mean primary liability for users who download and use such model weights in the UK – and since primary liability does not depend on knowledge, even an innocent purchaser of the downloadable model would be liable. That could lead to consequences not considered by this court. 

It is also worth mentioning that such a finding would only prevent the provision of downloadable, not cloud-based, versions of AI models. This is because the model weights are only supplied to users when a model is downloaded. For cloud-based versions of AI models, model weights are hosted on servers – there would be no infringement assuming those servers are based outside of the UK (more on this below). For this reason, AI developers will no doubt be ensuring their model weights are stored/hosted from outside of the UK.

What Getty did argue

The argument Getty did run hinged firmly on the wording of section 27 of the CDPA - that an article is an infringing copy if "its making in the [UK] would have constituted an infringement of the copyright in the work in question…". Getty argued that the making of the article (the model weights) involved the use of infringing copies (in the training dataset) such that the article is itself an 'infringing copy' (even though it does not and never has stored those copies). 

This is a 'but for' type argument - but for the copyright works existing in the training dataset (which the parties agreed they did), the model weights could not have been created. The judge rejected this argument on numerous grounds. In particular:

• Section 27 of the CDPA requires the article that is made to be a 'copy'. It is not concerned with a process which (while it may involve acts of infringement) ultimately produces an article which is not itself an infringing copy.
• Sections 22 and 23 of the CDPA are not addressed to the act of reproduction of a copy, but to the downstream dealings in an article which is 'an infringing copy'.
• The notion of an infringing copy includes within it the idea that there is a 'copy'.

In short, this is a finding that the secondary liability framework is not aimed at the scenario in question – that framework focuses on dealings with articles that are copies, not whether an article is made with the benefit of a copy. Looking at the wording of the legislation and its purpose, it is difficult to disagree with the judge's finding on this as a matter of statutory interpretation. 

Looked at this way, it was always going to be difficult for the judge to find liability. This is especially so given the UK's on-going policy discussions about the interaction between AI and copyright, including the extra-territorial reach of the UK's copyright laws. 

A finding of infringement would have had huge potential implications, meaning that there could be secondary infringement where any article is produced with the benefit (anywhere in the process) of an infringing work, subject to the knowledge and other requirements. Those downstream (eg those who import/possess a model) could also be liable if they have the requisite knowledge. 

If the UK wishes to go down that route, then it is for parliament to consider, especially given the complex and distinct nature of the UK's primary and secondary liability regimes.

Can an 'article' be intangible? 

The judge held that an 'article' can be something intangible. She gives two (related) explanations:

• Storage has developed to include intangible methods. The judge starts by noting that, at the time the CDPA was enacted, storage media capable of storing electronic copies "…would have involved physical, tangible means of storage such as a floppy disk…". She goes on to note that "…in today’s world, electronic copies can be stored in an intangible 'cloud' with no physical instantiation." In view of this, she concludes that parliament must have intended to cover any means of storing electronic copies, whether tangible or intangible. She finds support in section 17 of the CDPA, which says that copying includes "reproducing the work in any material form" including "storing the work in any medium by electronic means". While, on face value, the judge's ruling has some attraction, there is an obvious hole: the cloud is, in fact, a server somewhere; there is no such thing as 'intangible' storage. It would be more accurate to say that intangible works are stored on tangible servers and hardware and imported/distributed via electronic systems and cables ie that there is a transmission of the works. Clearly, the server/hardware has not been imported into the UK. When put like this, the judge's ruling feels like a mixing of the primary (communication to the public) and secondary (importing an article that is an infringing copy) regimes. While some may say, "So what?", it should be noted that the two regimes have different knowledge requirements and the overlap between them has been carefully considered by parliament (eg extending secondary liability for copyright infringement to intangible broadcasts and cable transmissions was expressly rejected by parliament when the CDPA was first enacted). That's not to say that the two regimes do not now need further thought given the way technology has developed. 
• The article is the infringing copy. The second explanation given by the judge is that the article can be the infringing copy; they can be one and the same thing. Therefore, if an 'infringing copy' can be something intangible (which section 17 says that it can), then so can the 'article'. The judge rejected Stability's argument that "…the requirement that the 'copy' is an 'article' in sections 22 and 23 is “…an additional requirement which means that the secondary infringement provisions can only apply to copies in tangible form”. She did so on the basis that the "…drafting neither suggests nor implies that in addition to being a copy, an infringing copy must also be something else…". Again, while there is some attraction to this finding, it leads to some uncomfortable questions. First, it makes the use of the word 'article' in the legislation otiose. It doesn't necessarily follow that what a copy can be in the context of 'infringing copy' should determine what a different term, 'article', should mean, especially when that term is expressed as being a different one to 'infringing copy'. Secondly, there appears to be some circular reasoning from the judge: there had to be an infringing copy in the article because otherwise the article wouldn't be an infringing copy.

Effectively, this is a finding that there can (in an appropriate case) be secondary liability for importing/dealing with an 'infringing copy' (without the need for there to be an article), which seems to go against the wording and structure of the CDPA. Some policy thinking influenced the judge's decision on this point, when she found that "if the word 'article' were construed as only covering tangible articles, this would deprive authors of protection in circumstances where the copy is itself electronic and it is then dealt with electronically". However, as noted above, in such circumstances primary infringement could have a role to play so authors may not necessarily be deprived of protection.

If the judge had found infringement…

It's worth pointing out that, even if the judge had found secondary infringement, it would only have applied to the downloadable versions of Stability's model, not the cloud-based versions. The model weights are only supplied to the UK when users download Stability's model; no model weights are supplied in the UK when users access the cloud-based version of the model (which are hosted on servers located outside of the UK). 

It would have left Stability (and other AI developers) free to supply the web-based versions of its model to UK users. Such a ruling would also have meant 'business as usual' for cloud-based SaaS providers hosted from outside of the UK. 

Lessons for dealing in copyright works

The court was also asked to consider whether Getty was entitled to bring its claim in respect of certain of the allegedly infringing works which had been exclusively licensed to it. Those licences were governed by New York law. 

There is a difference between UK and US copyright law on what constitutes an exclusive licence. Under UK law, there can only be one licensee (to the exclusion of everyone else, including the copyright owner) for a licence to be exclusive such that the licensee can benefit from the statutory powers to enforce the licence. However, under US law, an exclusive licence can be shared between one or more parties and still fall within the statutory definition. 

Which construction applies will depend on the applicable law and the interpretation of the agreement under that law. Here, a New York law governed licence granted 'exclusive' rights to the Getty group (being more than one entity) rather than a single entity (such as the holding company). As such, it did not qualify as an exclusive licence under the CDPA and Getty could not sue for infringement under it. 

This is an important point to check in any infringement action – with Getty being unable to sue themselves, it would have required the underlying photographer to have been a named claimant in the action (even if Getty were to have run the case for them). 

It is also something that should be factored in when dealing with multinational copyright exploitation. To achieve the same commercial effect of enabling a corporate group to exploit a work but retaining the exclusive licensee privileges under UK law, the licence should be granted exclusively to a single rights acquisition entity in the group, with that entity having the power to grant (non-exclusive) sub-licences across its group.

There was also a minor (in the context of the case) but nevertheless practically important decision on what can constitute a signature in writing for the purposes of the CDPA (in this case, for grant of an exclusive licence). The judge found that typing “I Agree” or clicking a button labelled “I Accept” is sufficient to constitute signing a licence. While this is common practice for copyright agreements entered into online, it will be welcome relief to those who seek to rely on them to give title to copyright works. Use of DocuSign (ie electronic signatures) was also taken to amount to a signature for these purposes. Nevertheless, it will be important for licensees and transferees to keep records of such acceptances (and of which specific terms were accepted, especially when the terms and contracting entities change over time). In some cases Getty did not have records of individual contributors who had typed "I Agree". As such, Getty could not prove they had sufficient title to the works of those contributors.

What the court did not decide

It's important to keep in mind that the two primary copyright infringement claims were dropped by Getty - for more details on why and the implications, see here. 

Had Getty been able to show that the training of Stability's model occurred in the UK, the outcome of the case might have been very different. This illustrates how important jurisdiction is to these claims. The issue in future English cases will be for claimants to find ways to show that the development and/or deployment of AI models in some way touches the UK jurisdiction eg through location of engineers and computing resources used for acquiring training datasets and carrying out training activities.

Likewise, there might have been a finding of output infringement in this case had the images Getty alleged were copied been more original/unique and had there been a clearer link between them and the alleged infringing outputs. Future cases may well be in that 'ball-park'.

Scope for an appeal

There is obvious scope for an appeal, but since it will not cover whether the model weights contain a copy of the copyright works given the parties' and their experts' agreement on this issue in this case (and will not cover primary liability either), it will not deliver all of the answers. Nonetheless, the decision does bring the scope of the UK's primary and secondary liability regimes into focus – which the government may (or may not) factor in when it delivers the results of the AI and copyright consultation. 

In this series:

• Getty v Stability: the trade mark ruling

• Testarossa: easier to prove genuine use of legacy brands?

• easyGroup v Premier Inn: Court rejects family of marks monopoly and clarifies section 10(3) protection
• Mercedes radiator grille found to be distinctive