The big British legal battle that is exploring the copyright obligations of AI companies, involving Getty Images and Stability AI, was back in court this week, with record industry trade group BPI intervening in support of the visual media company.
This case, now in the Court Of Appeal, centres on whether making an AI model available in the UK that was trained on unlicensed content in the US constitutes copyright infringement under UK law, given the actual unlicensed copying happened in a different country. Last year the High Court in London ruled that it does not, a ruling Getty is now appealing.
That ruling was based on the judge’s interpretation of so called ‘secondary infringement’ in UK copyright law. According to the BPI, that interpretation of the law - if left unchallenged - suggests AI companies can train their models “offshore” to avoid UK copyright rules and then still commercialise their models within the UK market. Which could have “significant consequences” for the music and wider creative industries.
However, even if that precedent did stand, it doesn’t mean every AI company could avoid copyright infringement claims in the UK by doing their training in another country.
While hoping the High Court’s ruling will be overturned on appeal, the BPI was also keen to stress that future copyright claims against AI companies in the UK might not swing on the same arguments as the Getty case, citing the German litigation pursued by collecting society GEMA against Open AI and Suno.
In the UK case, Getty accepted that Stability’s AI model didn’t actually contain copies of its images. However, in the German cases the courts ruled that the AI models did contain copies of GEMA-controlled works because of ‘memorisation’. If a UK court likewise concluded that any one AI model contained copies of works it was trained on through memorisation, then the Getty ruling wouldn’t be relevant.
Getty originally sued Stability AI for both primary and secondary copyright infringement, claiming that the AI firm trained its Stable Diffusion model by making unlicensed copies of images in the Getty image libraries. However, because Stability’s training didn't happen in the UK, and because it was agreed its AI model didn’t contain copies of Getty’s images, Getty was forced to drop the primary infringement claims.
That put the spotlight on the secondary infringement claim. Under UK law, a party can be liable for secondary infringement if they are shown to be importing into the UK, and then commercialising, “an article which is, and which he knows or has reason to believe is, an infringing copy of the work”.
That kind of secondary infringement claim would usually be made against a person or company that sold physical copyright infringing goods that were manufactured abroad. For example, if CDs are pressed without the copyright owner's permission in China and then imported for sale in the UK.
The copying didn’t happen in the UK, but that person is still liable for copyright infringement if the pressing of the CDs “would have constituted an infringement of the copyright in the work in question” under UK law.
The question for the high court was whether that rule - that traditionally applies to physical products imported into the UK - also applies to the Stable Diffusion AI model. The judge concluded that it did not.
Most people in the music industry - and the wider copyright industries - believe that a similar principle should apply to AI models trained abroad but commercialised in the UK, so that they must have been trained according to UK copyright law oblivious of where the training actually takes place.
Proposals have been made to amend UK copyright law to that effect, including by the All Party Parliamentary Group For Music, via a report it published on AI back in 2024.
But could the existing rules around imports and secondary infringement be interpreted in such a way that they also apply to AI models?
Getty and its allies in the copyright industries would like that to happen. Though lawyers for Stability obviously argue that shouldn’t happen, with a legal rep for the AI company stating that it’s not for the courts to extend copyright liabilities in that way, with only Parliament empowered to do so.
Meanwhile, as the BPI pointed out in its intervention into Getty v Stability, the key legal question would be different in any future copyright case against an AI model if it could be argued that the AI model itself contains copies of the works on which it was trained, as happened in the GEMA cases.
Harry Jewson from law firm Burges Salmon, who has been following the Getty v Stability case, writes that the “BPI has highlighted the significance of what it says is not in issue in the appeal, namely whether AI models generally contain copies of copyright works on which they are trained”.
While Getty accepted that the Stable Diffusion AI model does not contain copies of its images, the BPI cites the German GEMA cases where “the Munich Court has found that AI models did contain copyright works on which they had been trained, having ‘memorised’ them”.
So called ‘memorisation’, Jewson explains, is “a machine learning process whereby a trained model is able to reproduce significant parts of its training data as output”.
“On this basis”, he goes on, “BPI says the Getty judgment must be viewed as having ‘no wider significance’ on that particular issue, which is fact-specific to each AI model”.
A ruling on the Getty appeal is expected towards the end of this year or in early 2027. As with all the many AI copyright cases around the world, the ruling will be important to all creators and copyright owners, even though this particular case specifically centres on the copyright in images and photos.
Although, as the BPI has stressed, that doesn’t necessarily mean the final Getty ruling sets a legal precedent that is relevant in any future copyright case against an AI company in the UK.
