Yet another court filing in yet another Suno legal case, but this time Suno is responding to the lawsuit led by singer-songwriter Jason Isbell that accuses the music AI company of infringing the publicity rights of musicians rather than infringing the copyright in their recordings.
That’s based on the allegation that Suno’s AI model “indexes musicians by name”, so that it can generate new music “in the style” of existing artists, exploiting each musician’s “identifying attributes” - or ‘musical identity’ - for commercial gain without the consent of those musicians.
It’s a novel argument that makes this Suno lawsuit stand out from the pack. It also means Suno’s legal team can’t simply respond by cutting and pasting the same old fair use arguments they’ve employed in all of their other legal battles with the music industry, instead having to brush up on how state-level publicity rights work across the US.
Having done so, Suno’s lawyers have - somewhat unsurprisingly, given how important this is for Suno - reached the legally convenient conclusion that connecting musicians to musical styles within a back-end database does not infringe anyone’s publicity rights, and therefore they want the judge to promptly dismiss Isbell’s entire lawsuit.
The AI company’s lawyers write that - although publicity rights exist in most US states - no state provides a publicity right that “prohibits a musician in a studio from saying to their producer, ‘let’s go for something here with more of a Jason Isbell vibe’".
Providing of course what they then create doesn’t directly infringe the copyright in Isbell’s songs or recordings, and isn’t released in a way to suggest Isbell was involved.
“What Suno does is no different”, the lawyers then argue. “If a user asks Suno for a song in the style of Jason Isbell”, they continue, “Suno will effectively translate that request into one for a song in the style of ‘contemporary Americana singer-songwriter with finger-picked acoustic guitar’, and generate an accompanying output containing the ‘clear male vocals and country twang’ characteristic of that genre”.
And, Suno’s legal team insists, “there is nothing unlawful about that”.
Publicity rights - often called personality rights in other countries - are relevant when it comes to certain aspects of generative AI, for example when AI imitates the voice or likeness of a specific performer in the content it outputs, generating what are increasingly referred to as ‘digital replicas’.
Most people agree that AI-generated digital replicas exploit a performer’s publicity or personality rights, meaning AI platforms - and for that matter record labels that control an artist’s copyrights - must get consent from the artist before allowing any imitation of voice or likeness in that way.
Even when an artist’s voice or likeness isn’t imitated, if an AI tool encourages users to include specific artist names when prompting an AI - to generate music “in the style of” that artist - that also likely exploits the artist’s publicity rights. However, Suno is keen to stress in its new legal filing that it doesn’t do that.
It writes, “the Suno user interface does not affirmatively display artist names, or provide options to generate songs using a menu of particular artist names, or even suggest the idea of artist names in prompts in general”.
Indeed, technically users can’t even include an artist’s name when prompting Suno's AI. Although Suno does employ a sneaky trick when users do that, which was highlighted in Isbell’s lawsuit: basically it replaces the artist’s name with suggestions for styles and genres that describe that artist’s music.
Suno admits that that’s how things work in its legal filing, but stresses that Isbell hasn’t actually alleged “that any such description ever mentions an artist’s name, let alone does so to convey the impression that the output was actually the work of an artist or that an artist endorses the output or the platform”.
Therefore, it argues, the sneaky trick of swapping an artist’s name for suggested styles and genres “does not comprise a violation of the right of publicity”.
Whether or not this argument convinces the judge remains to be seen, but properly assessing Suno’s argument requires understanding a little about machine learning.
Isbell’s lawsuit claims that the fact Suno replaces an artist’s name with suggested styles and genres means that it must have connected artist names to musical identities in the back-end, and is now using that connection to make its genre and style suggestions. And in doing so it is basically providing an “in the style of” music generation service.
Suno insists that isn’t how it works - stressing that its AI is “not designed to output music that could be passed off as the work of an existing artist”, and that that principle is “baked into the very architecture of the platform”.
Suno “does not train its generative music AI model to recognise artist names at all”, its lawyers add, and “there is no ‘artist name’ metadata field in the source material used to train the model”.
However, Isbell’s original lawsuit alleges that Suno did in fact pair recordings with artist names when training its model. It says Suno “obtained audio and systematically paired it with the names of the musicians who created it”, and that this pairing “is what later enables Suno’s trained model to associate a specific name with a specific set of identifying attributes”.
But even if Suno’s description of its behind-the-scenes processes is accurate, that doesn’t automatically mean Isbell’s complaint isn’t valid. Just because the artist name does not appear in the source material does not necessarily mean that the artist name is not being used - albeit at arm’s length.
To understand why, it helps to know a little more about ‘features’ in machine learning and how they can be used to identify patterns in data.
Extracting features from music - such as tempo, key, genre and instruments - or, if you like “contemporary Americana singer-songwriter with finger-picked acoustic guitar” - helps turn music into data which can be used to match against a user’s text-based prompts - or an abstraction of those prompts - to identify the types of features they are specifying, which can then be used to inform the generative AI model that creates synthetic music.
Extracting features might include initial pre-processing steps including, for example, separating songs into stems to identify exact instruments and vocal tracks before extracting additional features from those stems, such as the vocal timbre, range and other qualities.
When this is applied to the songs of a particular artist, that process can build up a detailed mathematical representation of how that artist sounds, including genre and overall style - but also more granular qualities like how their voice sounds, and their ‘signature’ instrumentation.
Suno’s own legal filing confirms that the artist’s name does in fact play a role - just before the request is sent to its AI model. When a prompt includes an artist’s name, explain Suno’s lawyers, the model - rather than seeing that name - “sees the genre and style descriptions that a third-party tool has identified as being associated with that name”.
Type in ‘Duran Duran’, for example, and the name is replaced with something like “new wave, synth-pop, pop rock” - meaning, Suno says, that “the portion of the model that processes a user’s prompt never even sees an artist’s name”.
In other words, there is a system that maps artist names to descriptions of their music, and that sits separately from the source material used to train the model and is only used to map an artist name entered by a user to the features of that artist’s music.
So when a user enters a prompt, that “third party system” is queried, the mapping from the artist name to the features of that artist’s music is carried out, and the features identified are then used to inform the generation of the synthetic music in the style of that artist.
Whether that “third party tool” is an external service that Suno queries via an API, or whether it is an internal tool that is “third party” to its AI model, is not clear.
However, there are various datasets in existence - including Spotify’s audio features dataset - which can help with that process. Spotify restricted access to that data for new developers in November 2024, citing security, and further restricted developer access to its API platform earlier this year, saying that advances in automation and AI had changed the “risk profile” of providing that access to third parties.
How and where this mapping happens - and the fact that it sits outside the model itself - does not necessarily matter. Suno says that “the band name itself is not part of the direction that reaches the engine of creation”. But the artist’s name is still the input that determines which features are selected and used to direct the model - it has simply been moved one step away from the model.
The more important question is how detailed that mapping is. Suno’s lawyers say names are replaced with “broad genre-level descriptions” and the styles Suno shows users when it replaces a name are indeed broad - Isbell’s lawsuit says that, for Taylor Swift, Suno showed ‘pop’, ‘country’, ‘country pop’ and ‘synth-Pop’.
When the name of salsa singer Hector Lavoe was entered, Suno rejected it, but, says Isbell’s legal filing, “with no new prompt” it still produced a salsa track “imitating Mr Lavoe’s bright, naturally slightly nasal vocals”. That suggests that while Suno uses broad category terms in its user interface, a far richer and more detailed set of features - down to the qualities of an artist’s voice - are informing the generation itself.
Suno’s filing also makes two different arguments that sit somewhat awkwardly against each other. On one hand, it argues that it has not used anyone’s identity and that an artist’s name is only ever turned into “broad genre-level descriptions”, but then goes on to argue that the lawsuit is actually just a copyright claim in disguise and should be dismissed on that basis.
But within that argument it says that - on Isbell’s allegations - what an artist name “summons” is a sound, including “genre, instrumentation, vocal timbre and phrasing” that is “learned from plaintiffs’ recordings”.
Those distinctive qualities - the “bright, naturally slightly nasal vocals” and “vocal timbre and phrasing” that Suno has “learned” - are important, because they can be protected by publicity rights in several of the US states covered by Isbell’s lawsuit.
And courts have previously ruled that soundalike claims fall outside of copyright law, because a voice is not protected by copyright.
So while Suno is not conceding that its system works as Isbell claims, its lawyers do not dispute the musician’s account of the role an artist’s name plays in turning a prompt into an output.
And, in fact, they rely on that account to argue that the case is really just a copyright dispute, dressed up as a personality rights claim. “To the extent plaintiffs’ claims rest on the use of their names”, they write, “they seek to control a sonic style, and invoking those names adds no qualitatively different element to plaintiffs’ claims” - in other words, nothing that would take the case outside the scope of copyright law.
And, as anyone familiar with machine learning will understand, a good enough machine learning model does not need an actual name to make these sorts of matches - with enough data, and good enough feature extraction, machine learning can find - or inform - near-perfect matches.
So by taking an artist name, abstracting that artist name in the user interface into broad genres, but then “summoning” a set of features including “vocal timbre and phrasing” that has been “learned from recordings” of a particular artist and passing those features - without the artist name - to the AI model, it does not really matter whether or not the artist name is passed directly.
Whether the court will get into this level of detail on the technical side remains to be seen. Even if it does, it is still unclear if connecting an artist name to a musical identity in the back-end behind the scenes - however that might work - is an infringement of publicity rights. Unless, of course, you’re Suno’s legal team, in which case it definitely is not.
