Stability AI defeats Getty Images in landmark copyright case

The long-awaited judgment in Getty Images v Stability AI Limited has been handed down, seeing Stability AI Limited ("Stability") largely victorious and Getty Images ("Getty") calling for the UK government to take steps to support content creators.

The proceedings were issued in January 2023 and attracted global attention for tackling critical questions regarding the use of copyright works to train artificial intelligence (AI) models. While Getty abandoned its primary copyright infringement claims, the judgment still provides important guidance on secondary copyright infringement and the approach to how UK legislation applies to AI development.

As a reminder for those of you who may not have read our previous article, Stability's deep learning AI, named "Stable Diffusion", is alleged to have been trained on millions of images from Getty websites, for which Getty holds the copyright. Getty argued that Stability imported into the UK, otherwise than for private and domestic use, an article which is, and which it knew or had reason to believe was an infringing copy of Getty's copyright works (the "Copyright Works").

The outputs of Stable Diffusion were also claimed to infringe Getty's trademarks to the extent the synthetic images bear those trademarks in the form of watermarks.

This judgment sets a precedent for how courts may balance innovation with intellectual property rights, shaping the future of AI governance and compliance.

Copyright secondary infringement claim

Getty did not contend that Stable Diffusion itself is a copy, nor that it stores copies of the Copyright Works. Instead, Getty advanced its case of secondary infringement under sections 22 (importing infringing copy) and 23 (possessing or dealing with infringing copy) of the Copyright, Designs and Patents Act 1988 (CDPA) by stating that the making of Stable Diffusion's model weights (model parameters) would have constituted infringement had it been carried out in the UK. It is worth remembering that knowledge is a required component for secondary infringement.

The relevant sections of the CDPA state:

In order to make out its case for secondary infringement, Getty therefore had to establish that Stable Diffusion is capable of being an "article" (under ss 22 and 23) and an "infringing copy" (in accordance with s27). Mrs Justice Smith was explicit that "it is not possible to determine whether Stable Diffusion is capable of being an infringing copy without a clear understanding of what Stable Diffusion actually is". In so doing, Mrs Justice Smith laid the groundworks for a fact dependent case-by-case basis analysis being required in order to assess copyright infringement and AI. The question follows, what is Stable Diffusion?

It was agreed by the experts that Stable Diffusion does not itself store the data on which it was trained, and that once the network was trained, the process of generating image outputs could be run without the training data. In these circumstances there was no evidence of the model having "memorized" the Copyright Works, nor any evidence of any image having been derived from the Copyright Works.

Instead, Getty's case relied on it being enough that the making of the model weights would constitute infringement had the act been carried out within the UK. Turning to the interpretation of "article" and "infringing copy", and importantly, whether Stable Diffusion as interpreted by the experts above, is capable of being either.

Notably, there is no statutory definition of "article". Mrs Justice Smith considered that "article" must be construed with ss22 and 23 in mind. Specifically, any article which falls within the definition of an "infringing copy" must be considered capable of being subject to these secondary infringement claims. The court therefore had to consider whether an infringing copy encompasses intangible articles, as Getty claimed and Stability disputed.

In order to arrive at any conclusion as to the interpretation of "infringing copy" the court had to consider what is meant by copying, otherwise inconsistencies could arise from within the statute.

Turning back the pages in your statutory book you arrive at s17 (in particular s17(2)) of the CDPA which provides:

The Court held that in the modern world, where storage can occur in an intangible cloud with no physical instantiation, and given the breadth of the phrase "any medium", s17(2) must be capable of including intangible storage mediums. Accordingly, intangible assets must be similarly capable of being "an article". In arriving at this conclusion, Mrs Justice Smith dismissed Stability's arguments that the various acts of secondary infringement and concepts of "importation" and "possession" only apply where there is a tangible article.

A further test remained. Stability argued and Mrs Justice Smith agreed that, regardless of whether an article is tangible or intangible, an infringing copy must be a copy. Mrs Justice Smith stated "the essence of the infringement is that there has been an infringement of copyright by the reproduction of the work (including by storage in any medium by electronic means) in any material form...an article becomes an infringing copy when the act of reproduction occurs. From that moment the article is an infringing copy – but it ceases to be an infringing copy once it no longer contains the copy."

Whilst Stability accepted that transient copying was sufficient, it argued that an infringing copy must at some point contain a copy. Mrs Justice Smith agreed and held that an AI model which derives results from or uses a training process involving exposure to copyright works is not itself an infringing copy. Whilst the model is altered during training by exposure to the copyright works, it does not store any of those works.

The model weights are not an infringing copy and do not store an infringing copy - "[t]hey are purely the product of the patterns and features which they have learnt over time during the training process". Consequently, Stable Diffusion fell outside the CDPA for secondary infringement as it did not constitute an "infringing copy".

Trademark Infringement

Getty were, in part, successful in their arguments that images produced by Stability's AI models, which included a likeness of their watermark, had infringed their trademarks. However, the judgment on trademark infringement was limited in scope – as the court assessed the output of each model and version of Stable Diffusion in turn, rather than as a collective. As a result, only three instances of infringement were established.

Where the court ruled that Stability had infringed Getty's trademarks, they found that there was sufficient likelihood of confusion and association between the AI created watermarked images and the original Getty images used to train the models. This was because, although users of the model would know that the image outputs were not Getty originals, they might be led to believe that the model was trained on Getty images under a licensing agreement between Stability and Getty.

Getty's arguments that the replications of their watermark gave Stability an unfair advantage and could cause detriment to their reputation if used on unfavourable or illegal imagery both failed, due to lack of evidence.

What have we learned from the judgment?

The judgment will be seen as a victory for AI, but it has confirmed that AI can fall within the provisions of secondary copyright infringement if certain criteria are met. While the application of primary copyright infringement provisions to AI remains unanswered for now, Mrs Justice Smith also considered what her findings would have been, had she found that Stable Diffusion fell within the interpretation of "infringing copy", and she concluded that:

• Making available for download intangible copies onto a local device in the UK is sufficient to meet the definition of "importation". The act of importation does not have to be carried out by the defendant.
• Where a hosted service is provided from servers outside the UK, merely gaining access to the hosted service does not involve importation – regardless of interpretation "there could never be any act of secondary infringement by reason of the provision of remote software services".
• Stability knew or had reason to believe that the training data had been scraped from the internet without consent and in respect of works which are protected by copyright.

The case and judgment have been against the backdrop of the UK government's ongoing debate as to how to accommodate AI within existing and potential legal and regulatory frameworks.

Last month saw the government launch its proposed AI Growth Lab, testing grounds where regulations would be relaxed for specific periods under supervision. The government has called for views on the proposals and established working groups to consider the balance between the interests of content creators and AI developers.

The debate on copyright and AI will rage on, both at government level and in the courts. As cliché as it sounds, it really is a case of watch this space!

If you'd like to understand how this judgment might affect your business or wish to discuss AI and IP law further, speak to Kate Swaine or any member of the team for bespoke, expert legal advice.