Copyright through the lens of Getty Images v Stability AI

Summary

This article provides an outline of the key copyright implications of the first English judgment on generative AI.  The judgment has been eagerly awaited by the creative and technology industries, as well as the government which, so far, has been unsuccessful in resolving the tension between them arising from the use of copyright works by AI developers.  While the decision did not deliver a decisive victory to either party, it has revived the discussion around secondary copyright infringement in the digital age.

The High Court judgment only briefly ended a three-year legal battle between Getty Images (Getty) and Stability AI (Stability) in the UK.  Both parties have been granted a leave to appeal, Getty on copyright, and Stability on trade marks.  Furthermore, Getty’s US case against Stability continues, alongside other high profile cases against generative AI.  

Notwithstanding the appeal, the case is the first attempt to sets out a legal framework and judicial thinking on primary and secondary copyright infringement of works used for AI training and synthetic output.

Background

Getty is a leading image library licensing photographs and video internationally.  Stability is a company within a corporate group developing and making available generative artificial intelligence models.

The development of generative AI software relies on processing of large volumes of human-created copyright works (in Getty’s case, photographs scraped from its website) to train predictive capabilities of data processing software.

The processing of data in the course of training informs the value, or weight, given to each piece of data (e.g. processing of an image of a black cat will assign a higher value to data readable as “black”). This results in an AI model with highly sophisticated order of weights.  Stability’s model, Stable Diffusion, and the model weights are made available either as a piece of software to AI developers or to users through an online platform.  

Copyright claim

Getty alleged that images available on Getty’s website were used by Stability to train the models thus giving rise to:

1. Primary infringement of its copyright under s. 16 and 17 CDPA
   
   • during the training stages (the Training Claim); and
   • through the creation of synthetic output which is substantially similar to Getty’s photographs (the Output Claim):
2. Secondary infringement, with the Stable Diffusion model weights being an “article” imported into the UK without Getty’s permission and with knowledge that model weights were an “infringing copy” of Getty’s works under ss.22, 23 and 27 CDPA (the Importation Claim).

As the case progressed to closing submissions, Getty was forced to abandon the Training and the Output Claims due to evidential difficulties centering around proving that copying took place in the UK.  In the absence of the Training Claim, the Importation Claim became the main battleground.  With evidence of importation unchallenged, both parties agreed that model weights were imported into the UK and the decision would be a matter of statutory interpretation.

Judgment

The court decided in favour of Stability and held that model weights were not an infringing copy and, accordingly, that there was no secondary infringement.

The deciding consideration in the judge’s decision that model weights were not an infringing copy was the fact that model weights do not store any copyright works, but instead ‘…learn the statistics of patterns which are associated with certain concepts found in… the training data…’. Accordingly, it was concluded that “Stable Diffusion does not store or reproduce any copyright works and nor has it ever done so.”

Getty submitted that the wording of s. 27 CDPA does not require copies of works to be stored in the imported article; the section merely requires that the making of the article must constitute a copyright infringement if the article was made in the UK.  Getty’s position was that the CDPA must be interpreted in accordance with the “always speaking” principle of interpretation, allowing a statute to be adapted to new forms of uses of copyright works.  The judge agreed but gave greater emphasis to the natural meaning of the words “infringing” and “copy”. The judge pointed to the meaning of “copying” in s. 17 CDPA which requires reproducing the work and concluded that this meaning is instructive in interpreting section 27.  In the judge’s view, an article is an infringing copy if the manner in which it is made requires it to reproduce copyright works.

Consequently, the court decided in favour of Stability and held that model weights were not an infringing copy. However, in case this interpretation was wrong, the judge discussed the facts of the case if the Court of Appeal were to arrive at a different interpretation (paragraphs 603-609 of the judgment).  

What are the practical implications of this case?

The course of litigation in this case and the judgment demonstrate the difficulty in enforcing copyright where works are used to train artificial intelligence (AI) models and in particular the, so-called, generative AI models.  Both parties are large international corporations, able to withstand the rigours and costs of litigation.  The litigation history included an application for summary judgment, an expert list, and battles around the scope of disclosure and technology capable of interrogating Stability’s gigantic unstructured data sets.  The judgement spreads over 200 pages and addresses novel points of copyright law which would be daunting to even the most sophisticated of litigants.  While our reliance on AI increases day by day, enforcement against AI companies is likely to be prohibitive for all but the wealthiest creators.

At the substantive level, the case points to a statutory lacuna relating to liability for secondary infringement under ss. 22, 23 and 27 of the Copyright, Designs and Patents Act 1988 (CDPA). When enacted, these sections addressed situations in which copyright works were unlawfully copied out of jurisdiction onto a medium such as a book or a CD-ROM, and the “pirated” copies were brought into the UK. With electronic delivery and streaming making physical carriers largely redundant, litigation focus shifted to the right of communication to the public. The Getty case confirms that intangible digital articles created outside the UK and offered to UK clients may, in certain circumstances, give rise to an “importer” liability. The judge’s interpretation of the CDPA, however, does not consider that derivative digital products – such as AI models – are “infringing copies” in the meaning of s.27 of the CDPA. This would appear to undermine the protection of UK copyright owners and shield technology companies using their works outside the jurisdiction.