AI faces the music

A host of record labels, including Warner, Sony Music and Atlantic Records, filed suit against and Suno and Udio, in a pair of cases complaining about the “convincing imitation” of tracks by generative AI. According to pre-litigation correspondence, the defence of these companies, as in other US cases relating to Generative AI will hinge on the assertion that any copying carried out was fair use.

How Suno and Udio allegedly hit the wrong note

The allegations against Suno (which has an agreement to be included in Microsoft’s Copilot AI chatbot) and Udio are that they copy large numbers of recordings, clean these up to make them suitable for use as training data, then derive parameters that form the AI model from this training data. For example, when a user prompts the service with a text input (e.g. make a jazz song about New York), the service generates one or several tracks, usually with lyrics, on the subject requested.

In a well trained  AI model, the aim is usually for the output to be based on a generalisation of the corpus of training data. However, in this case it is suggested that the Suno and Udio services suffer from the problem of ‘overfitting‘, in which an output is likely to closely resemble parts of the training data. This is not the basis of the case: it is alleged that even if the model was improved, this would simply hide the copyright infringement which went into writing it. However, this overfitting provides the record companies with what is in their opinion a smoking gun: they “betray that the model was trained on the copyrighted recordings”.

Imitation or innovation?

The record companies have clearly spent some time playing with the models.

Suno’s model was tasked with prompts such as “1950s Rock and roll, jerry lee lewis, sun studio” which produced a song with the very nearly familiar title “You shake my nerves and you rattle my br”  (that isn’t a typo, and you can make your own joke about AI being half a brain). This song contains the line “Goodness gracious, great balls of fire”, and yes, there is a ‘large vocal leap’ to the word ‘great’. There are number of other examples: the lyrics of Sway by Michael Bublé were near exactly reproduced by the prompt “Canadian smooth male singer 2004 jazz pop buble sway latin mambo minor key”. A simple prompt of “70s pop” produced a song called “Prancing Queen” – you can make your own guess at which Swedish band the track brings to mind.

Udio appear to have some basic protections against attempts to directly generate material which sounds similar to an input track as, in some cases, the prompts used by the record labels have had to avoid use of the artists names. However, “my tempting 1964 girl smokey sing hitsville soul pop” generated a song not too dissimilar to ‘My Girl’ by the Temptations. Some of the prompts are flattering to the artists: “m a r i a h c a r e y […] holiday […] remarkable vocal range” produced a song actually called ‘All I want for Christmas is You”, in which the ersatz singer doesn’t care about the presents under the Christmas tree. Rhyming also seems get around the security provision: the prompt “[…] song about queens that dance, by a Swedish band that rhymes with fabba […] from an album that rhymes with ‘jarrival’” produced a song called, in this case, “Prancing Monach”. For those who enjoy rhyming, Truce Stringbean, Cohnny Jash and boldplay also appear in sample prompts set out in the pleadings. It is an interesting side point to consider whether the fact that copyright protection mechanisms were included by Udio – however easily circumvented – will be in their favour (hey, at least they tried!) or a tacit admission of their system’s ability to produce close copies of songs.

Pre-empting fair use, the pleadings set out that, in the view of the record labels, the similarities, the sheer volume of material, and the potential impact on the business models of the Record companies rules this out.

You can listen to some tracks generated based on these prompts here. They are impressive, and the power of the technology is clear. Yet they are also worse than the originals. It is hard to imagine, for example, an ABBA fan genuinely seeking out Prancing Monarch to get their fix, but it is also true that they could play a role in ‘elevator music’ scenarios. Half heard, they carry with them the familiarity of the original.

What’s next in the legal battle over AI-generated music?

It is, of course, relatively early days in the development of these cases. However, listening to some of the outputs, it seems hard to believe that a court could conclude that running training data through an AI engine data can somehow cleanse it of its copyright, allowing substantively similar recordings to be produced. That said, the internet has a long history of scraping:  search engines would not work without it, for example. Therefore, banning web scraping at this stage would seem impractical.

The practical solution may perhaps be found in opt-outs – i.e. , allowing a copyright holder to instruct that its material should not be ingested as training data. This seems to be the way forward in Europe, featuring the AI act (although as yet exceedingly cumbersome in application). Alternatively, it could be that the ‘overfitting’ is indeed deemed to be the issue, and that so long as the output is far enough from any particular track ingested as training data, AI music generation based on copyrighted material can continue. A further option may be good old-fashioned licensing – could the model which killed Napster but gave us Spotify work again for licensing training data?

Whatever the case, we’ll be keeping an ear out for progress.