Fake Drake
In mid-April of 2023, fans of Canadian superstar rapper Drake broke out in jubilation over a new release by their idol as the song “Heart on My Sleeve” appeared on Spotify and then on YouTube, after which it went viral on TikTok. It soon became clear, however, that what seemed to be a duet between Drake and his Canadian compatriot The Weeknd was actually an AI-generated deepfake perpetrated by an anonymous Internet user known as “ghostwriter977”, pictured sporting a sheet over his head and sunglasses.1
This “fake Drake” not only attracted worldwide media attention but also alarmed Universal Music Group (UMG), the world’s largest music conglomerate, whose sublabel Republic has both Drake and The Weeknd under contract. In a public statement, UMG expressed its disapproval of this and related practices, claiming violation of its copyrights.2
Do AI Applications Violate Copyright?
With that, the discussion of whether AI violates copyright had reached the music industry. When attempting to assess whether AI tools actually do so, we first need to look more closely at how generative AI applications work.
Artificial intelligence is referred to as generative if it is capable of giving rise to new content independently. To do so, inputs—audio files, for example—are fed into an AI system, where they are turned into so-called hidden layers by way of statistical and mathematical processes. Modern AI systems interconnect multiple thousands of such hidden layers. This involves complex transformations of the input data known as deep learning, which is something that even an AI system’s programmers cannot fully comprehend. By way of providing an illustrative example: WaveNet, conceived by the Google subsidiary DeepMind, is a self-learning AI algorithm that can learn to create music by trying out new combinations, which initially requires several months and an extreme amount of computation time. Once the learning process has been initiated, however, the AI system becomes capable of producing results at ever-shorter intervals. The amounts of data processed as part of this are immense. Every musical note that WaveNet generates requires 16,000 micro-music samples per second from the data set.3 This makes it impossible to determine just which input data the AI system has used and linked together in order to arrive at a given output. From a technical standpoint, what happens in this AI application are not duplication processes but instead complex transformations of data leading to a new piece of music that can no longer be understood as a copy of pre-existing musical works.
Even so, the importation of input data—insofar as these are protected by copyright—does represent a copyright-relevant duplication process. It is this fact that has been at the centre of copyright infringement lawsuits by the majors in the USA—against the AI company Anthropic4 (in which Amazon has invested billions) on the music publishing front and against the Internet-based AI music generators Suno5 and Udio6 where the music label business is concerned. The major labels accuse these AI companies of having used their copyright-protected content without permission and demand that they pay to license their music catalogues. The AI businesses and the tech companies that back them have countered by claiming that training AI models is covered by US law’s “fair use doctrine” and that permission to use the data in question is hence unnecessary, as one can read in an October 2023 response by Google to a US Copyright Office request for comments.7
In the states of the EU, where “fair use” is not anchored in copyright law, the EU Database Directive states that the retrieval and further processing of data can be disallowed during the copyright term of 15 years but explicitly permits use for the purpose of text and data mining.8 This is contradicted, however, by the EU’s recently passed regulation on artificial intelligence (the AI Act) in which rights owners are conceded the ability to opt out of text and data mining.9 This requires the expenditure of quite some effort, though, since it must first be ascertained which AI company is using the data to begin with. It will most probably be left to the courts to determine once and for all whether and in what form AI companies are infringing on copyright in the realm of music.
Economic Implications
The concrete motivations behind the conflict between the music conglomerates and tech giants are first and foremost economic in nature. The streaming economy has entailed a massive increase in the value of music catalogues because streaming services like Spotify must license both the publishing as well as the master rights to musical recordings. This is also the reason for the current boom on the music rights market in recent years, with spectacular catalogue purchases by the majors as well as new protagonists backed by big finance. Over the past five years, private equity groups like Blackstone and wealth management giants like BlackRock have ploughed a total of USD 13 billion into the acquisition of music rights in hopes of enhanced returns for their investors.
If the rights owners succeed in forcing the AI companies to pay for the use of their music catalogues for AI training, they will have tapped a new source of revenue in the music industry that will serve to further boost these music catalogues’ value. The tech companies, for their part, naturally have no interest in paying for the use of music catalogues, since that would drastically increase the cost of developing their AI tools. However the courts ultimately rule, there will inevitably be economic consequences where the use of music rights is concerned.
