Comments on the potential copyright claim by the BPI over AI deepfakes

1. How does this differ from a situation like the Getty Images case, considering as the AI model training is done (predominantly) by users rather than the website itself?

"Jammable may try to argue in its defence that it benefits from “safe harbour” rules, which generally protect online platforms from infringement claims on their platform where they are not aware of the infringement – Jammable currently has a published Digital Millennium Copyright Act policy on its website.

"There is, however, nuance to the protection provided by “safe harbour”. If a platform intervenes (e.g., by manually reviewing uploaded content, so that it has specific knowledge of infringement) or plays an active role to give it knowledge of or control over the content uploaded to its platform, it may be liable for the communication to the public of infringing material.

"Whether Jammable’s use of user-provided content (which is then manipulated by their AI model and made available more generally by their “community uploaded voices” tool) would negate "safe harbour" protections remains to be seen. It’s a complicated area which is very fact specific. For example, compare the results of the Tuneln Inc case in 2019 and the ECJ’s decision in the Cyando case in 2021. Here, the interesting nuance is that the infringing content uploaded by the user is not directly being made available to the public by Jammable on the service – there is a technical intervention by the AI system and the output is an AI model trained on potentially infringing content. It’s not clear whether Jammable stores or uses content uploaded by its users in the back-end of the model."  

2. Any differences in the legality of training AI using existing songs and recordings fed into the system by users rather than the broad datasets used by other generative AI systems (the EU offers data scraping protections for the latter)?

"The reliance on Text and Data Mining (TDM) permissions varies across jurisdictions – notably the BPI’s complaint will be based under UK law, where the current form of the TDM permissions would likely not apply (unlike the EU’s more permissive opt-out regime, the UK’s TDM permission is limited to non-commercial research).

"Existing songs and recordings will benefit from various forms of copyright, almost always in this context including rights in the sound recording (which for most signed artists will be controlled by their label). There will be other rights holders too, with rights over the copyright created by the songwriter in the music itself being protectable separately. Typically these rights holders, in particular the major labels, are experienced in protecting their rights. As such, if the use by Jammable is found to be infringing, there are potentially many interested rights holders that suitable licenses would need to be obtained from."

3. As well as copyright infringement, do issues of breach of data or privacy laws, misrepresentation protections or reputational damage enter into play given the ability to mimic an existing figure's voice?

"There’s no right to a person’s voice per se in the UK, but there are risks to both Jammable and end-users here. If someone’s voice (or what they are saying in a recording) is capable of identifying them, it might constitute personal data. In this case, Jammable would owe the person various obligations under UK GDPR, including an obligation to inform them of its processing of their personal data and provide them with privacy information.

"End users of Jammable’s service might not realise that by using AI-generated voices online, depending on the use-case, they could be engaging with false-endorsement or passing off principles (for example, if using the fake voice to promote goods or services without being suitably transparent) and might engage more serious issues under defamation or fraud if the user is using the fake-voice to cause harm to the person whose voice they’re mimicking."

4. Any other thoughts on the topic?

"This is a new and evolving market, and technology providers are pushing the boundaries of what is acceptable. Some are more open about this than others – OpenAI is on record as saying it is impossible to train AI without using copyrighted works. Other key players are now negotiating licensing deals for data sets (for example, Reddit has now reportedly licensed the rights to train AI on the user-generated content on its site). Ultimately generative AI models are only as good as the training data underpinning them, so I expect to see more and more licensing deals, and the development of industry norms, over the coming months and years."

Andrew's comments were published in Law360, 11 April 2024.