Ludovic Hunter-Tilney
That’s the mellow sound of “Dust on the Wind”, a song by the band the Velvet Sundown. They play classic rock in the style of 1970s bands like Kansas and The Eagles, it’s all very FM radio. But retooled for the age of streaming and playlists. The Velvet Sundown’s debut album Floating on Echoes came out in June.
It racked up a million streams on Spotify in less than a month, catapulting the band to fame. But people noticed something strange about this quartet of retro rockers. Velvet Sundown were oddly elusive. There were no interviews with them, no tour dates, no website, and there was something even more peculiar about their publicity photos. Four airbrushed Californian dudes with eerily perfect faces and identical hands holding hamburgers in a celebratory meal as though they had never seen a griddled patty before. So who exactly are Velvet Sundown? The answer is no one, or at least not the people they claimed to be.
Audio clip
It turns out Velvet Sundown and their hit songs are totally AI generated.
Audio clip
The Velvet Sundown is a band that gets more than a million monthly listens. Turns out their music is generated by AI.
Audio clip
The band changed its profile on Spotify to say the project is guided by human creative direction and then composed, voiced, and visualised with the support of artificial intelligence.
[AI MUSIC PLAYING]
Ludovic Hunter-Tilney
The Velvet Sundown was an elaborate deception. Everything from the band’s lyrics to the melodies to the artwork were made by generative artificial intelligence and their success on streaming platforms has thrown up a lot of questions. How did they get so many streams? Why weren’t they flagged as AI by music platforms? And how did the Velvet Sundown music get so precise, so convincing that they were able to pass as a real band in the first place?
This is Tech Tonic from the Financial Times. I’m Ludovic Hunter-Tilney, the FT’s pop critic, and in this two part series we are looking at how AI is impacting the music industry. In the last episode, we explored how AI technology has evolved, whether it has the potential to take people’s jobs. And if it could present a threat to the very practice of music making itself.
But AI music generators may have an Achilles heel. They’re trained on literally millions of songs harvested from the internet and campaigners, record labels and artists say that includes stealing songs that should be protected by copyright law. The AI companies behind the music generators are coming up against an age old music industry adage, where there’s a hit, there’s a writ.
This is part two: Theft machines.
Every new piece of music contains traces of other pieces of music that came before it, but AI song generators take this recycling process to an extreme. The Velvet Sundown are a case in point. Their sound accurately mimics an analogue era of vinyl and guitars, long hair, and a thick fog of marijuana smoke. The song you were just listening to, “Dust on the Wind”, has the vibe of Kansas hit “Dust in the Wind”.
And that feeling of familiarity isn’t just because of the almost indistinguishable titles. It also comes from the way these AI models learn.
Ed Newton-Rex
To train today’s AI models in a given domain, you need tons of examples of that thing to learn from. So you need tons of real music.
Ludovic Hunter-Tilney
That’s Ed Newton-Rex, who you heard in the previous episode. He’s a composer and the CEO of Fairly Trained, a non-profit that advocates for ethical AI models, but he worked in generative AI for more than a decade.
Ed Newton-Rex
So AI companies, they either go and send out their own web crawler that will kind of crawl tons and tons of websites, find links to music files, ultimately they’ll scrape them, they’ll download them, or they will go and get access to some other library that someone else has already done that for. And then essentially, the way these training algorithms work is these models try to recreate what they hear, right? So you’ll stop the music at like five seconds and you’ll try to get it to replicate kind of what comes next.
You can think of it like this and if the next kind of sounded outputs is the next one in the original song, you essentially give it a good mark, right? And you do this just untold numbers of times with untold numbers of pieces of music. So what this does is it builds up this model that has this understanding of music such that it can then essentially generate its own music.
Ludovic Hunter-Tilney
This learning, according to Newton-Rex is on a totally unprecedented scale.
Is it at all analogous to the sort of processes that you would do as a composer if you were to go and just, let’s say, listen to lots of music by Benjamin Britten and then go and compose something which was clearly influenced by it. Is this really like a sort of computer equivalent at what it’s capable of being able to assimilate and reproduce?
Ed Newton-Rex
It’s really not. The ways that these processes work are totally different. As a simple example, basically to train one of these models, you need billions of data points. That is not how we learn music.
I do not go and consume the entire canon of Western music and learn its intricacies and copy it repeatedly and then output something fresh that by the way, every so might quote something I’ve learned, like that’s a simple example of how these processes are totally different.
Ludovic Hunter-Tilney
The result is you’ve got these models that have mined the internet for millions of songs.
For Newton-Rex and other musicians, the problem is not just that these models are so powerful, it’s that they may have been trained on those musicians’ music without their knowledge, without their permission.
Ed Newton-Rex
Given that this music belongs to other people, is being taken without their permission, is being copied huge numbers of times in this process, which is ultimately where copyright’s infringed. What is happening here is basically theft, dressed up as competition. You know, you have this model at the end and the AI company will say, you know, it’s just like a human competing with you. But actually it’s very, very different ’cause they have, they’ve stolen this music.
Ludovic Hunter-Tilney
Newton-Rex says that it wasn’t always like this. Back in 2010 when he started working in generative AI, AI companies stuck to training models on music that either wasn’t copyrighted. Or that they’d paid for.
Ed Newton-Rex
But then in 2022, across the whole AI industry, you suddenly had this moment where what had previously been happening behind closed doors, where people have been sort of researching what would happen if you just took more work to train on without asking.
Suddenly people started releasing these models commercially, which was unheard of. And at that point, the whole AI industry almost overnight shifted and you got this huge shift towards basing these models and what’s essentially theft, you know, it’s going and just gathering whatever work you can to train your models on and ignoring the rights of the people who own that work.
Ludovic Hunter-Tilney
To be clear, AI companies don’t deny that they’ve trained their models on vast swaths of music gathered from the internet. Suno, one of the most advanced AI generators, has said that training includes all musical files of reasonable quality that are available on the internet.
It also acknowledges that lots of the music available is copyrighted, but both Suno and Udio, another leading AI generator, say their use of this music to train their models would fall under an exception to copyright law called ‘fair use’. And ‘fair use’ has a pretty broad definition, allowing for copyrighted material to be used in circumstances like criticism, commentary, education, or in some cases, for there to be some creative and transformative use.
The idea that AI companies should be allowed to use copyrighted music seemingly with such ease is causing controversy. Suno and Udio are both facing lawsuits from copyright holders in the US and Europe. And in the UK, the government is considering taking things one step further.
Ed Newton-Rex
Currently, an AI company taking your work and training on it in this way in the UK is illegal. The government wants to legalise it. The government is actively trying to legalise the theft of people’s creative labour.
Ludovic Hunter-Tilney
What the British government is proposing is an amendment to the law that would actually give AI companies special rights to use copyrighted music. We got in touch with a government spokesperson who rejected the idea that they were legalising ‘theft’ as Newton-Rex describes it and added that, quote, no decisions have been taken, but our focus will always be on enhancing the ability of rights holders to be compensated and paid fairly while enabling access to high quality material to train leading AI models in the UK.
I first came across Newton-Rex’s name earlier this year because he put together an album I reviewed. It featured a surreally impressive super group. Han Zimmer, Damon Albarn, Annie Lennox, Kate Bush, The Pet Shop Boys, and about a thousand other artists. Here’s a snippet of what it sounded like.
[SILENCE]
Yes, that is indeed the sound of silence. The album called Is This What We Want is a protest album, a silent protest album. It was designed to illustrate what the artists say will happen if proposed changes to UK copyright law go ahead. Empty studios, muted musicians, the end of songs sung by the human voice. And this isn’t just a UK issue. A lot of governments around the world are eager to accommodate AI. No one wants to miss out on this revolutionary technology.
Ed Newton-Rex
I think in general, lots of governments right now feel a lot of pressure. I live in the US and the standard line here is, you know, we gotta compete with China. And you hear that in the UK as well. This idea that like, there’s AI race and if we don’t allow AI companies to steal all the musician’s work, we will somehow lose the AI race. But I think this is, this lacks so much nuance. I mean, you know, firstly you have AI companies talking about curing cancer and AI solving disease and making scientific discoveries.
And let’s be clear, they haven’t done any of this yet. But ultimately, even if you think that that’s possible, you don’t need Ed Sheeran songs if you want to cure cancer. Like all respect to Ed Sheeran, you only need people’s music if you want to build AI models that will compete with them.
Ludovic Hunter-Tilney
Copyright infringement is complicated. And while AI music companies might currently claim ‘fair use’ in training their models, they’re cautious when it comes to reproducing real artists. When I tried out Udio in the first episode of this series, it seemed quite effective at producing songs in a certain style. Death metal say, or reggaeton. But when I asked it to copy specific artists, it seemed to misfire. Take the Britpop song I asked it to make in the style of Oasis.
[AI MUSIC PLAYING]
It sounded nothing like Oasis. It did sound a bit like Guns N’ Roses. This happened a lot with Udio who say they don’t generate artist likeness without permission. If I asked it to copy the style of a specific artist, it would remove the name from the prompt and often make something that sounded completely different.
Like when I asked it to make a song in the style of Madonna. Song like “Ray of Light” by Madonna is what I’m asking it.
[AI MUSIC PLAYING]
So Udio has really botched this one because it’s created what sounds to me like a really bad copy of a Kylie track. It’s got nothing to do with Madonna’s “Ray of Light” there. It’s even quoting, spinning around. You got me spinning round and round, which I mean that has to be a little bit like, I don’t want to get myself mired in all of the thickets of legal tangles that are taking place now, but a bit cheeky really.
It looked like it was time to talk to a real lawyer. I met up with Gregor Pryor, an entertainment lawyer at the firm, Reed Smith. He’s been in the industry for more than two decades and has been involved in some landmark disputes around music.
Gregor Pryor
I think about generative AI and what’s happening right now as the next big legal frontier for the music industry.
Ludovic Hunter-Tilney
In the US last year, the world’s biggest record labels, Warner Records, Sony Music and Universal Music Group launched a legal battle against Suno and Udio claiming copyright infringement on an almost unimaginable scale.
Gregor Pryor
So the music industry comes together, particularly on the recorded music side, and they say, right, OK, this isn’t on, and they join together to sue two companies, Suno and Udio, early stage start-ups, innovators, disrupters, you might argue, and the music industry says, hey, you are not allowed, or you should not be allowed to copy my music and store it and use it for the purposes of training your machine to create new music.
Ludovic Hunter-Tilney
This is not, for what it’s worth, the only case currently levelled against Udio and Suno. Earlier this year, the aptly named country musician, Tony Justice, launched a lawsuit representing the interests of independent artists. And there’s been a class-action lawsuit against Suno specifically, from the German copyright society, GEMA.
The music industry is well versed in fighting tech innovation. After the CD boom of the 1990s, the music business was riding high. Then suddenly at the end of the decade, file sharing services like Napster began proliferating.
Gregor Pryor
And then all of a sudden, wow, you can get your music anywhere for free from some unknown actor. And so the music industry embarked on a big campaign of litigation, a campaign for the public’s attention and sympathy for this bad behaviour.
Ludovic Hunter-Tilney
The record companies successfully sued file sharing platforms like Napster and LimeWire, but it was like a game of whack-a-mole. Other platforms popped up to replace them. Eventually, the labels ended up doing deals with streaming companies like Spotify to put their catalogues online.
Gregor Pryor
Spotify is what changed the game where they said to the music industry, well, your music is available for free everywhere now. What if I charge people a certain amount per month? That was the big sea change, and all of a sudden you were able to get all the world’s music for 10 bucks a month. That’s a huge bargain.
Ludovic Hunter-Tilney
That was a long time ago. Two decades on the balance of power between the music industry and Big Tech is much more one-sided. Today’s tech industry is bigger and more powerful than it’s ever been.
Gregor Pryor
The music industry, in relative terms, is incredibly small. It tends to be very cohesive and powerful because it is newsworthy. Captures the public’s attention. Everyone identifies with music, so it’s got a lot of other strengths, but in relative terms, it’s absolutely dwarfed by the tech industries. So it is a David and Goliath star battle.
Ludovic Hunter-Tilney
So can the music industry strike another deal? Suno and Udio are now reported to be in talks with the labels.
Gregor Pryor
The narrative at the moment is that the companies are negotiating. I think there’s a fear on both sides. You know, it’s a win or lose case. Either the courts say ‘it’s permitted’, in which case the record labels have lost and that’s very bad news for them. Or they say ‘it isn’t permitted’ and Suno and Udio losing the investors lose a lot of their money. So because it’s such a febrile and difficult case, a win or lose case, then the parties are inclined to negotiate. . .
Ludovic Hunter-Tilney
which leads to settlements . . .
Gregor Pryor
Or a licence, or a licence and most likely, and a licence if things go the way that I suspect they will go, the record labels will take some money in return for this training that they haven’t permitted, and they’ll license these services moving forward.
Ludovic Hunter-Tilney
You can imagine a world in which record labels in the AI industry come to some kind of arrangement where the AI companies get to keep training better and better models, and the labels, in theory at least, get a slice of that pie. They can make a bit of money off the AI boom. But where does this leave the artists?
Even if a deal is reached, it still won’t remove their big fear that they’ll be competing with AI music generators for listeners.
[AI MUSIC PLAYING]
I want to go back to the Velvet Sundown, who you heard at the beginning of this episode. They’re the AI-generated band posing as a groovy seventies soft rock outfit. Their songs accrued millions of streams and turned up in Spotify’s Discover Weekly playlists. All without any indication that the band was AI and the Velvet Sundown is far from the only example of AI music on streaming platforms.
It’s impossible to know the exact numbers. Deezer — one of the smaller platforms — says it’s dealing with 20,000 AI uploads every day. A recent report by music industry analysts Luminate puts the figure for the whole streaming industry at nearly 100,000 AI songs daily. And this is already starting to impact artists.
Hana Stretton
I had WhatsApp messages, I had emails, lots of Instagram messages saying, I think something’s gone wrong with your Spotify. There seems to be someone who sounds like Phil Collins singing on your page with a terrible, terrible artwork. Is this you?
Ludovic Hunter-Tilney
Hana Stretton is a singer songwriter based in Australia. She’s had a successful career on streaming platforms. Some of her songs have gone viral, like this 2016 track “The Thrill Of Loneliness”, which has been streamed 17.6 million times on Spotify.
Ludovic Hunter-Tilney
Stretton makes most of her income from streaming predominantly Spotify, and she spent most of this year working on a new album. But in August, something strange happened. She woke up to see that a song she had never heard of had landed on her profile alongside accompanying artwork of an airbrushed guitar lying on a bed of poppies all listed under her name.
[‘WHERE D’YOU GO’ CLIP PLAYING]
This is the song. Tt’s called “Where D’You Go?” And I think you’ll agree. It sounds pretty different from Hannah’s usual music.
Hana Stretton
I mean, I don’t, I don’t think it sounds anything like me. I think they’ve used Phil Collins. It sounds like him. It sounds like his voice. It’s quite a generic sort of seventies sound.
Ludovic Hunter-Tilney
It turns out that the song is AI-generated and had been uploaded without her knowledge by someone who had managed to overwrite her profile.
It’s not clear exactly why they did this, but there have been similar instances of streaming fraud reported. With bad actors attempting to hijack artists’ revenues. Concerned about both her reputation and her revenue stream, Stretton tried to get the song taken down, contacting both Spotify and the company who appeared to have made it, but it remained on her profile for a whole week. During that time, the income from those streams was going to someone else, and she was getting more messages from bemused fans.
Hana Stretton
‘Did you hit your head?’ is one that I got (laughter). Did you hit your head, this is terrible, or just dot, dot, dot. What’s going on? We waited three years for an album and this is what we’re getting.
Ludovic Hunter-Tilney
Eventually, Stretton was able to get the song removed, but she says this is happening all the time.
Hana Stretton
This is such a widespread problem. This is actually not the first time this has happened to me. There’s a reason why someone like me is targeted. I do have very good streaming numbers for someone who doesn’t have any promo, and I’m independent.
So I don’t have a label or a team behind me, you know, with lawyers coming out of my ears to be able to really protect myself. And I think it’s just upsetting if it tarnishes your good name. You know, I’ve been working for 10 years to build this independent career, and it’s very sad to have it so easily taken away.
Ludovic Hunter-Tilney
There are lots of problems that arise as a result of AI not being labelled, and it isn’t just limited to hijacking artists’ profiles. AI- generated music might just appear in people’s playlists without warning, and even where music is identified as AI, streaming platforms such as Spotify don’t necessarily take it down.
Liz Pelly
There are like tens of thousands of generative AI tracks delivered to streaming services every day. I think that a lot of actors in the music business who are hoping to capitalise on the wave of generative AI content are really just hoping that people won’t notice it.
Ludovic Hunter-Tilney
That’s music journalist Liz Pelly. She spent years looking into Spotify’s business model and is the author of Mood Machine. The Rise of Spotify and the Costs of the Perfect Playlist. Pelly says that Spotify doesn’t seem to do much to stop AI from appearing on its platform or even being pushed to listeners. Take the example of the Velvet Sundown.
Liz Pelly
One piece of the puzzle of how those tracks sort of like went viral on streaming had to do with the fact that they were appearing in Discover Weekly playlists on Spotify, so algorithmic personalised playlists.
Ludovic Hunter-Tilney
Spotify is the world’s biggest streaming platform with almost 700 million users. And it works by pushing algorithmic recommendations. It’s not just a repository of the world’s music. It tries to tailor music to your taste so that you can sit back and listen continuously, like listening to your favourite radio station. Its Discover Weekly playlists are a huge part of how the platform works. Listeners have streamed songs featured on these playlists more than 100 billion times.
Liz Pelly
A lot of those algorithmic recommendations are based on listening patterns across the platform. And I think the Velvet Sundown song, it had been added to like a third party playlist called Vietnam war songs or something like that, that had, you know, hundreds of thousands of followers and clearly they had some sort of strategy for juicing these tracks and algorithmic recommendations. And then what happens is like if certain tracks that fit a user’s taste profile are gaining lots of traction, it’ll boost the chances that it might end up in one of their algorithmic recommendations.
Ludovic Hunter-Tilney
Spotify, unlike some other streamers, does not flag when content is AI generated. And it doesn’t take AI music down unless it violates its terms and conditions like impersonating an existing artist.
We asked Spotify about this and they didn’t get back to us. We also spoke to Deezer, a French streaming platform, who told us that tagging AI tracks and excluding them from playlists isn’t impossible. They say they are so far the only platform that does this, but by comparison to Spotify, Deezer is really small, with about 9 million subscribers. And in the meantime, more and more AI music is appearing across streaming platforms generally, and the burden is mostly on us, the listener, to figure out whether or not the song is made by a machine or by a human.
Liz Pelly
Even now, something that I experience is this dynamic of people sharing a link with me and saying, is this AI, is this a real artist? Is this, do you know if this is made to generative AI? And I just don’t feel like people should have those questions. Like that almost is like, right now feels more urgent to me, is like making sure that people have the tools to know when they’re being served degenerative AI track.
Ludovic Hunter-Tilney
A big anxiety about AI music is that it will start seeping into daily life, slowly saturating the musical landscape, like the figurative frog being boiled, we won’t notice until it’s too late. At least if we’re told what’s AI and what’s not, then we can decide for ourselves whether we want to listen to it. And that might be good for artists too: if we listeners have an informed choice, then maybe we’ll decide we actually want real human-made music no matter how accomplished the AI music becomes. But will we?
Holly Herndon
I think the public will determine what they want. So if people decide that they really want music that was written by writers camps rather than algorithms, then that is what people will listen to.
Ludovic Hunter-Tilney
Holly Herndon is the pioneering musician I spoke to in part one of this series. She’s been experimenting with generative AI for a decade, and she says so much of popular music is already a production line, so why not automate it?
Holly Herndon
Often on the biggest hits that we’ll hear, you know, people will go to a writer’s camp and there will be 50 different writing credits on a single song. So in a way, it’s already a kind of collective intelligence, perfectly crafting the song to meet the conditions of the invisible algorithm that Spotify is setting.
So there’s already so much manipulation to the craft there that I see no issue in that being kind of fully automated. And I also just don’t think that all musics have to follow the same logic. I think it’s OK to have subcultures. You know, different music can have different functions in our world.
Ludovic Hunter-Tilney
What function will AI music fulfil in the future? I began this series feeling sceptical about claims that AI could seriously threaten the music industry. I’m less sceptical now.
While I remain doubtful about dystopian fantasies of the charts becoming flooded by AI songs. I am convinced that AI will increasingly play an everyday part in the making of music, either as aid or rival.
But this whole process has also made me reconsider the role music plays in our lives. It is the art for most closely and also mysteriously linked with our emotions. And the deeper I’ve delved into the world of AI, the more I found myself thinking about human-made music, what it is and what we value in it.
I went back to Dan Tarbuck. You’ll remember him from part one. He’s the composer and musician whose studio I went to visit. His company makes high quality production music. In response to the threat of AI, Tarbuck has decided to double down on the thing that he believes makes music so distinctive: it’s humanness.
Dan Tarbuck
You know, we make some really authentic things there where people can bounce ideas off each other, and we are really capturing that, and that’s what we’re really leaning into now. Human-made music and the most human possible way, where everyone’s improvising and just making things up on the spot. There’s mistakes in it. You might not even notice them, but you’d notice if they weren’t there. That’s part of the charm. It’s never gonna be perfect. Whereas AI, I think maybe is trying to be.
Ludovic Hunter-Tilney
So you were really are taking a bet then that the desire for flesh and blood musicians creating things in a room such as this, through moments of inspiration, which can’t be predicted is going to continue to be very important.
Dan Tarbuck
I think it’d be foolish to think that it’s not gonna affect us and that it isn’t affecting us, and I think it’s gonna affect every single industry, but we have to adapt and be able to offer something that we think is valuable, and hopefully other people see the value in that.
Music will always be a human thing at heart. You could even argue that AI music isn’t really music because it might sound like it on the surface, but there’s no meaning behind it. I just hope that people will always find the value in something which has been made by someone with feelings and with a story to tell, which at this point, AI doesn’t have.
Ludovic Hunter-Tilney
Thank you for listening to Tech Tonic with me, Ludovic Hunter-Tilney. The producers are Lulu Smyth and Josh Gabert-Doyon. The senior producer is Edwin Lane. The executive producer is Flo Phillips. Sound designed by Breen Turner and Samantha Giovinco. Original — not AI music — by Metaphor Music. Manuela Saragosa is the FT’s acting co-head of audio. Our fact checkers are Megan Hill and Miatta Mbriwa.
Read more on Financial Times News
