Why Getty’s appeal won’t address all UK AI copyright questions

The media company’s appeal, being heard on Wednesday 11 November, raises important points of law, but content creators and AI developers alike should not expect the eventual judgment to provide answers to all the questions they might have on how UK copyright law applies to generative-AI systems.

Below, we explain why, and look at the wider context in which the hearing is taking place, as other courts around the world grapple with how to interpret existing law on AI and copyright and policymakers continue to consider whether the law should be updated.

Primary copyright infringement questions will linger

Getty believes Stability AI trained its AI system Stable Diffusion using its copyrighted images without permission to do so. The problem it faced before the High Court was that it could not show that the training itself took place in the UK. This was fatal to its claims of primary UK copyright infringement.

Under the UK’s Copyright, Designs and Patents Act (CDPA) 1988, claims of primary infringement can encapsulate a range of infringing acts, including unauthorised copying or communication of copyright works to the public. For primary infringement claims to be made out under the CDPA, however, the alleged acts of infringement must have taken place in the UK.

Before the High Court, Stability AI challenged the primary infringement claims, including on the basis that none of its training of Stable Diffusion took place in the UK. Frustrated by an absence of evidence to prove otherwise, Getty withdrew its primary infringement claims during the trial and focused its attention instead on pursuing claims against Stability AI under other legal grounds.

The wider consequence of that is that, while content creators might assume claims of primary infringement would succeed in other cases against AI developers if their model training was undertaken in the UK, there has not yet been confirmation of that from the courts. It’s an omission that will not be addressed when the judges rule on Getty’s appeal and means there will be continued uncertainty for rights holders on this point.

The secondary infringement claims – an elephant in the room?

Getty’s UK copyright case did not end with the withdrawal of its primary infringement claims. The company also argued Stability AI was responsible for secondary copyright infringement too. It is these claims that form the centre of Getty’s appeal.

While the provisions on primary infringement address the direct acts of copyright infringement, the secondary infringement provisions under the CDPA address the acts that follow, providing scope for rights holders to pursue redress against those that import, possess, market, sell or distribute an ‘article’ featuring an ‘infringing copy’ of their copyright works.

The High Court’s ruling provided some guidance on how the secondary infringement provisions apply in an AI age.

First, the High Court rejected Stability AI’s argument that the concept of an ‘article’ only extended to tangible objects. For rights holders, this is helpful clarification that the secondary infringement provisions can apply to software such as AI and not just to infringing copies subsisting in books or other physical media.

However, while Getty satisfied one of the legal tests to make its secondary infringement claims, it was considered to fail on another.

The judge, Mrs Justice Joanna Smith DBE, said the concepts of an ‘article’ and an ‘infringing copy’ could not be divorced from one another. As a result, she considered that, for a secondary infringement claim to succeed, it would have to be shown that an ‘article’ itself constituted an ‘infringing copy’. She determined that this was not the case with Stable Diffusion. This is because, in her view, an AI model would not be considered an ‘infringing copy’ unless it has at some point contained a copy – permanent or transient – of the copyright works used to train it. 

The problem for Getty was that, earlier in the trial, it had accepted, as an agreed fact for the purposes of the litigation, that Stable Diffusion does not itself store the data on which it was trained. That fact was agreed between the parties after expert evidence was adduced. Because there was no storage of the training data on the AI system, it followed that there could be no ‘infringing copy’ of Getty’s images considered to subsist within it.

The High Court’s findings go to the heart of the broad question of how exactly gen-AI systems generate output when they respond to user prompts. Not all experts are agreed on the answer. For instance, academics at Stanford University said their research had found it is “feasible to extract large quantities of in-copyright training data from production LLMs” and cited earlier research that suggest this is because training data is memorised by the models and encoded in the models’ weights.

These differing views have led to different findings in other AI-related copyright disputes.

For example, the Regional Court of Munich backed music rights licensing body GEMA in two copyright cases raised against AI providers in recent months. In doing so, the court considered that copyright works used in the course of training the AI systems in question would continue to exist within the AI model by means of systematic memorisation.

An adviser to the EU’s highest court recently challenged the idea that data AI models are trained on is stored in or memorised by the model. He suggested instead that where copyright works are replicated in their entirety in the output an AI model generates, it could be down to the model being over-exposed to the same information within the training dataset and reproducing it based on a combination of its learning and statistical probability.

However, in a sign that the Munich court’s view is proving attractive to other aggrieved rights holders, illustrator Axel Scheffler, the man behind the images in The Gruffalo books, is reportedly considering bringing AI-related copyright claims in Germany.

Yet with the parties in the Getty case having agreed that Stable Diffusion does not store the data it was trained on, it seems highly likely that rights holders and AI developers will have to wait for a court in the UK to give its view on the concept of memorisation and the question of how gen-AI systems operate. This is because there are only very limited circumstances in which facts agreed by the parties before the first-instance court can be re-litigated on appeal.

Instead, the Court of Appeal’s consideration of Getty’s appeal on secondary infringement is likely to be limited to whether Mrs Justice Joanna Smith DBE correctly applied the relevant law in reaching her decision.

Before the High Court, Getty considered that its concession on a point of fact did not fatally undermine its secondary infringement claims, as it put forward an alternative interpretation of the legislation to the one the judge favoured. It argued secondary infringement would be demonstrated simply if training Stable Diffusion would have amounted to copyright infringement had it taken place in the UK. The company is expected to re-raise the matter as a question of law in its appeal.

Pressure on UK policymakers

Some commentators believe the High Court’s ruling highlighted a potential gap in the UK’s copyright regime when it comes to AI. This very assertion was made in a report published by the Institute for Ethics in AI at the University of Oxford  (36-page / 1.76MB PDF) in October.

The report, which focused on “the threats facing the UK’s creative sector and the policy solutions required to address them”, called on the UK government to make a range of changes to the UK’s copyright framework, including reforms to the primary and secondary copyright infringement provisions “to ensure that they adequately address the ways in which copyright works are used in the development of AI models and products”.

The government was also urged to introduce a new “legally binding” licensing regime, under which rights holders could “express permissions and preferences, and refuse consent” to use of their copyright works. It was further called on to “explicitly rule out text and data mining exceptions, automatic opt-out, and commercial research exceptions which do not uphold these standards”. All “products, models or services imported into or made available in the UK” would need to have “all relevant licences and consent to use UK copyright work in the UK market”, regardless of where that content was scraped, according to the report.

The proposed new framework would be underpinned by transparency obligations on AI providers, a series of new offences to support the enforcement of rights, and the designation of a regulator, which would be responsible “for regulating and enforcing the legal standards of ownership and consent, and transparency, traceability and disclosure”.

However, the government has previously said it does not favour an express licensing requirement, fearing that it would make the UK “significantly less competitive compared to other jurisdictions – such as the EU and US” in respect of encouraging AI development.

In March, though, the government confirmed it would not pursue other options for reform it had consulted on either, which included a regime built on transparency over use of copyright works and opt-out controls for rights holders. Instead, it said it would undertake some further evidence gathering and continue to monitor developments in other jurisdictions before deciding on next steps.

Not everyone is happy with the status quo position. In addition to facing the University of Oxford report’s call for reform, the government is subject to growing pressure from parliamentarians to act. Conservative MP Damian Hinds has brought a private members’ bill before the UK parliament with a view to requiring online publishers to be given greater visibility over the scraping of their content by online bots. A second reading of The Automated Online Software (Access and Transparency) Bill is due to take place in the House of Commons on 27 November.

It is not only domestic pressures that are falling on the government.

The US Court of Appeals recently ruled that unauthorised use of copyright works for the purposes of training a legal research search engine that used AI was not covered by the ‘fair use’ limitation in US copyright law. Viewed alongside the Regional Court of Munich rulings and contrasted with the first-instance decision in the Getty case, some may now argue that copyright law as it stands in the UK, in the AI context, is more disadvantageous towards rights holders than in the US and Germany, if not the wider EU also.

To date, the government has resisted the pressure to act on AI and copyright because it does not want to press ahead with reforms that are out of step with the rest of the world. Now, though, its critics will point to the status quo as itself putting the UK at odds with the emerging legal position being determined by the courts in other major jurisdictions. The appeal in the Getty Images v Stability AI case seems unlikely to change this.

The government has said it intends to pursue international consensus on AI issues when it assumes the presidency of the G20 in January. However, with much of the current discussion on AI focused on security, it is unclear whether copyright matters will make the agenda, far less be agreed upon by global leaders.