AI Music: The Infinite Jukebox

0:03Ludovic Hunter-Tilney: So I've tapped in create a song in the style of Oasis. Create a song in the style of Oasis. Let's have a listen.

0:21Ludovic Hunter-Tilney: That was absolutely horrific. That was like a bizarre Guns N' Roses. That was nothing like Oasis. Let's go for some pop rap and see what happens here. Let's go for pop rap about partying.

0:45Ludovic Hunter-Tilney: That was actually really quite good. A trashy, admittedly, but very upbeat and enjoyable pop rap song, which I would be only too happy to play at a party of my own. That is a winner. I'm Ludovic Hunter-Tilney. I'm the FT's pop critic. I spend my time covering gigs and critiquing albums by the biggest names in music. But today, I'm switching roles from reviewer to songwriter with the help of an artificially intelligent music generator. Let's do a death metal song with an indie pop chorus.

1:29Ludovic Hunter-Tilney: Not bad, actually. I thought they managed to do quite a decent death metal song. What I'm using here is a platform called UDO, one of many AI music generators that have recently emerged. Just type in a few key words, and in seconds it creates a brand new song, one that's never been heard before. Zero musical expertise needed, and you can ask it to make pretty much anything. Let's go to now reggaeton. And let's have reggaeton with a male vocalist singing about the office. Because there's not very many reggaeton songs that I'm conscious of which are about offices. Lots about clubs, beaches, having fun.

2:33Ludovic Hunter-Tilney: Well, that's quite brilliant. I, for one, would be really happy to roll in the office with my crew to that song. As you can tell, none of my AI-assisted efforts are liable to trouble the charts. For users like me, AI music generators are gimmicks, a fun way to make your own pop song. But in the last couple of years, generative AI technology has advanced dramatically. And these tools can now produce songs that are almost indistinguishable from man-made music. For the music industry, the storm clouds are gathering. The fear is that AI generators could flood the world with AI content, drowning out the work of humans, stealing the jobs of musicians, and even bringing an end to the process of music-making as we know it.

3:25Ludovic Hunter-Tilney: Is all this alarm just a case of technophobia? Or is the music industry really facing AI annihilation?

3:40This is Techtonic from the Financial Times. I'm Ludovic Hunter-Tilney, and welcome to the first in a two-part series on the impact of AI on the music industry. Part one, The Infinite Jukebox.

3:57Ludovic Hunter-Tilney: I'm a pop critic. I spend every day listening to and thinking about music. I'm fascinated by the way technology shapes how music is made. And right now, that means artificial intelligence, particularly generative AI. The idea of AI learning from data sets and then producing something entirely new. It's sending shockwaves through all industries and music is no exception. I'm not yet sure how I feel about it. I guess you could say I'm an AI agnostic, which is why I'm making this series, to find out more. So I started by talking to one of the most influential musicians working in AI today. Someone who's really enthusiastic about its capabilities.

4:43There's this idea of contributing to something that's greater than oneself. And that's kind of how I see these AI models. I see them as a grand collective human accomplishment. Holly Herndon is a musician and artist who's based in Berlin. For the last decade, she's been building her own generative AI models, training them on music until they create new and interesting sounds. I kind of started working with machine learning before the most recent wave of generative AI models. And so the first question that you're kind of asking yourself when you're trying to train a model is what's the data set?

5:17And as a composer, it felt like a creative choice, like I needed to make my own data set. Herndon grew up in Tennessee, in the American South, where she was immersed in the choral traditions of the church. So when it came to training her early AI models, she wanted to work with choirs. So that's what we did. We put together an ensemble in Berlin and we met about once a week and we recorded singing sessions. And then we began to train our AI, which at the time we called Spawn. We were really deliberate about creating bespoke training material for our models. And because of that, the models sound very unique to us rather than the kind of generalized models that most people are using today.

6:03The culmination of those early experiments with her generative AI model Spawn was an album Herndon put out in 2019 called Proto. Listening to some of the songs, you can almost hear Spawn learning as it absorbs the sounds of the choir. In the track Canaan, you can hear the ensemble singing a sample of the data being fed into the model.

6:31And then in this track, Evening Shades, listen to Spawn trying to mimic the human singing.

6:48Finally, in the track Godmother, the model creates its own sounds. It's almost like a primitive cry symbolic of an early attempt at AI music generation

7:04Lately, Herndon's AI models have evolved even further

7:13Last year, for an exhibition she and her collaborator Matt Dryhurst staged at London's Serpentine Gallery she went on a tour of the UK recording choirs in church halls across the country. We wrote a songbook for AI training, so if you sing through all the songs in the songbook, all of the phonemes and sounds of the English language would be covered enough for a machine to be able to understand the voice of that ensemble. And then we recorded 15 different choirs singing from the songbook. We travelled all the way up to Scotland and through Wales and really covered a broad range of vocalists.

7:50and it was a wild trip I mean at the time my son was only one years old so I packed him in a car with my mom and we drove around the UK recording these choirs with my little one-year-old in the back seat and then all of that music was put into a collective data set and then we created a kind of supermodel that we were then able to prompt and create compositions for the exhibition

8:38Ludovic Hunter-Tilney: Using generative AI to make music used to be the preserve of innovators like Herndon, who built their own models. But now there are dozens of easily accessible AI models that can make any kind of music you want at the touch of a button. Platforms like Udeo, which you heard me experimenting with at the top of the episode, and Suno, another powerful AI model, can produce entire songs that sound uncannily like the work of real people. These mass market models work in much the same way as Herndon's. They learn the rules of musical composition by being fed all types of music. The difference lies in how the source material is obtained.

9:19Ludovic Hunter-Tilney: Herndon's models are trained on the choral music she's gathered with the consent of the choirs, while models like Yudio and Suno consume vast swathes of music harvested from across the internet, likely without the consent of the musicians that produced it. The result? Incredibly sophisticated models that have essentially learned from the entire back catalogue of popular music, offering up the possibility to produce anything you can think of. This is completely unprecedented. Nothing in the history of music remotely compares to it. And just as early AI did for Herndon, these new, publicly available models offer up the opportunity to make music in exciting new ways.

10:02I'd always been searching for years and years and years, wondering when the next form of sound synthesis was going to pop up. And then suddenly someone plays me something they made on Udio, and it was what sounded like a pretty high-quality disco record. And that felt like a huge jump. Tom Furse is a founding member of British indie band The Horrors. They started as a gothic-tinged garage punk band in the early 2000s before embracing a more experimental electronic sound. first place synthesizers and keyboards, and he's always been inspired by music technology. He's specifically interested in sampling, the art of using pieces of existing music to create something new.

10:46For him, the latest AI generators are less about churning out whole AI songs and more a useful tool for generating samples. It's much more practical, to be honest with you. It's just like, I would love to have a vocal sample that said these words. and there's now a way of me doing that and getting a variety of those samples out.

11:06Ludovic Hunter-Tilney: Take me through how that would work. You're making a piece of music and you want to have a vocal sample. How would you use an AI platform to go and create that? So it's probably worth spelling out the traditional way. This is really common. You know, a producer, they might hear in another record a little bit of vocal, a little snatch of vocal. They think, there's something good there. That's something really interesting. And they incorporate that into their production. Or, you know, you could work with a session singer and, you know, get a singer into your studio and be like, I want you to sing these words in this way.

11:33In an AI context, I can go to a platform like Udio and it lets you type in your own lyrics. So I can type in the lyrics and then describe the style, the kind of genre that I imagine that vocal to have. I can take the vocal out of the generated pieces of music that it's made. So I can get 12 different vocal samples all singing the same thing with different singers and different voices and different styles. and that artistically is really interesting.

12:03Ludovic Hunter-Tilney: So you've produced the voice sample with its melody and then you would be incorporating that into a piece of music which you would be writing yourself. So you're sort of riffing off the melodic phrase that you've generated. Yes, I'm incorporating it. I'm probably not taking like a whole verse or a whole chorus. We're talking about half a second of sound, a word here or there, stitching together maybe an ambiguous sounding phrase. There's even a vowel sound pitched around, sped up and down, placed in the right way, used maybe as a percussive element as opposed to a melodic element. There's all kinds of ways of approaching this.

12:37Tom Furse uses these AI tools to inspire and enhance his own compositions. He's still in the creative driving seat. But what happens if the AI, capable of creating whole songs on its own, takes over from the human creator completely? Could it start competing with human musicians and eventually replace them altogether?

13:23is really hard. And Vista's Robert Smith. As a market dampened, guess what? We were able to monetize 18 companies. And of course, I'll take you behind the scenes of Bain's latest research, including our newly released mid-year private equity report. Subscribe to Dry Powder today.

13:42I began noticing the mainstream emergence of AI music a couple of years ago. Back in 2023, a song called Heart on My Sleeve went viral on TikTok and YouTube.

13:59It featured two of the biggest stars in music, the rapper Drake and singer The Weeknd. But it turned out it wasn't Drake or The Weeknd. Both their voices had in fact been AI-generated by a TikTok user called Ghostwriter977.

14:19At the time, there were predictions that AI-generated voice clones would sweep through music. Two years on, that shows no signs of happening. AI voice generators are popular for creating TikTok spoofs of favourite artists,

14:33Ludovic Hunter-Tilney: but their use remains that of a gimmick, a toy. Will the same happen with AI-generated songs? So far, the results have still been at the gimmicky end of the spectrum. Last year, a parody Schlager song made with UDO by an Austrian DJ called Butterbro reached the giddy heights of number 48 in the German charts.

15:02Ludovic Hunter-Tilney: Earlier this year, a band called The Velvet Sundown racked up over 750 ,000 Spotify listeners, even though it was revealed that the music and the photos of the band itself were generated by artificial intelligence.

15:23But this was a prank rather than a chart-quaking eruption of AI classic rock. The millions of streams that Velvet Sundown have notched up pales into insignificance next to the billions of pop's biggest names. For me, as a pop critic, these examples of AI-generated music are a joke. But for many musicians, they are the first warnings of a potential catastrophe. AI will be able to create music that is essentially indistinguishable from most human music. It kind of already can. That's musician and composer Ed Newton-Rex. He used to work for an AI music company himself, but not anymore. Now he campaigns for musicians' rights in the face of the threat of AI.

16:07Yuval Nohr-Harari recently said that the coming AI wave, you can think of the impact it'll have on people's jobs as like a billion immigrants. And I think in music, this coming wave is basically like a billion music college grads suddenly competing with you. Probably in those billion, you're not going to get the next JS Bach, right? Like AI is not going to be the next Bach. It's not going to be the next Paul McCartney. But what it is going to do is it's going to make it incredibly hard to make any money at all in this industry because you're going to be competing with these incredibly cheap models and they're offering them pretty much for free.

16:46Ludovic Hunter-Tilney: Would you describe AI in that case? I mean, you make that sound like an existential threat to the very practice of music itself. I think it is. Absolutely. I mean, I do think AI music is an existential threat to the practice of music. Of course, it puts people out of work. We're already seeing this. You have film directors who have already said they're going to use AI music going forward in all of their products. AI music is all over Spotify. It's all over all of the streaming platforms. You know, 18 % of daily uploads to Deezer, the sort of French alternative to Spotify, 18 % of uploads every day are AI generated.

17:20You can assume it's the same number or greater on Spotify. AI music is everywhere. It's competing with humans already. And you've seen this in art already, where I know artists who have had to give up on their art careers because of AI art. You see it in writing, where when ChatGPT was launched, almost overnight, freelance writing tasks on freelancer platforms fell by almost 10%. The effect of generative AI is predictable. It's common sense, right? Like if you have a hyperscalable, very cheap alternative to the human creator, it's going to reduce demand for those human creators. That is what's happening in music.

17:59Pop music and technology have always been inextricably linked. Recorded music began when Thomas Edison invented the phonograph in 1877. He initially thought it would be a kind of musical toy. Then, in 1925, the first microphone was introduced to the music studio, ushering in the era of electrically recorded music. You can imagine the alarm felt by jobbing musicians fearing that people would prefer to listen to their music at home rather than pay to go see it live. One newspaper at the time worried about the capitulation of the world's leading musical artists to the power of the microphone. More recently, there have been examples like the drum machine.

18:41In 1980, the Japanese company Roland released the TR-808. It was a small keyboard-sized box that allowed musicians to program complex drum rhythms without the need of an expensive drum kit, or, crucially, a drummer.

18:57Ludovic Hunter-Tilney: And by the mid-1980s, the drum machine was redefining pop music. Listen to the programmed drumming, featured on Marvin Gaye's classic soul song, Sexual Healing, from 1982.

19:15or the mechanical beats in the background of Whitney Houston's 1987 hit I Wanna Dance With Somebody.

19:28The 808 may have threatened to reduce the demand for session drummers in the 1980s but ultimately we still have drummers today as well as all the exciting new music that the drum machine helped musicians create. My point is that pop music is always evolving, and it's always absorbed new technologies to its advantage. So why can't the same be true for generative AI?

19:54Its critics say that AI is just a whole different beast, with abilities on a completely different scale. AI music generators are capable of not just replacing experienced musicians and their instruments. They can replace the composers and songwriters too. Unlike a drum machine or a microphone or even a phonograph, AI has the capacity to take over the entire creative process. And some are worried this might spell the end for whole sections of the music industry. This is the control room. Got our brand new SSL Origin 16-channel desk.

20:30Ludovic Hunter-Tilney: So that's the mixing desk. Yes, yes. Some lovely microphones and a really nice-sounding live room. This is Dan Tarbuck, the head of production for Bleach Music, a company that makes what is known as library music. The themes, instrumentals and jingles used in films and TV shows, radio adverts and yes, podcasts like this one. He gets session musicians in to record in his studio. It's called The Crypt, based in an old church in London. Now it's such an inspirational room for any musician who comes in here. They always comment on how much they love it because there are dozens of amazing instruments around.

21:04So you've got a 50s Deegan vibraphone over there. Really beautiful instrument. Hammond organ.

21:15Very, very classic kind of stuff you hear in all sorts of old soul tunes and things. We've got 110-year-old piano. Some really old drum kits. Guitars, amps through there. Beautiful whirly here, which is a very rare instrument.

21:30Ludovic Hunter-Tilney: It looks like a sort of wooden little piano. Yeah, I think it's walnut.

21:42Ludovic Hunter-Tilney: Library music is an important but often underappreciated part of the music industry, worth more than a billion dollars, and employing musicians, composers and producers from all over the world. Tarbuck reckons nothing beats the sound of musicians jamming together in the same space. We did a Boogie Woogie album with some really top players.

22:05They literally made everything up on the spot with really old instruments, so it's our piano, which gets progressively out of tune as it goes, which is great for the music. It sounds great and it sounds authentic and the people doing it are the real deal. Yeah, so, I mean you can listen to that all day, so good. Yeah, it's irresistible. Um, we've got, uh, this one was really cool. So this album was called New Jazz Grooves and again it was largely improvised in the moment. He's an amazing bassist, unreal. And then the drummer here, his groove is all over the place in the best possible way. It's so wonky and so cool.

22:42A bit of a sort of guerrilla's fairground thing going on there. Yeah, I can see what you mean. It's not following a set pattern, it's just changing based on how the drummer's feeling and how everyone's grooving together, which is something humans are very good at.

22:56Ludovic Hunter-Tilney: Humans might be very good at it, but it's easy to see how library music could be overtaken by AI. If an AI music generator can create all the music you need for your TV program or film and give it to you cheaply or even for free, why would you opt for the human-made version? I've spoken to a lot of people about this. Some people are thinking, oh, it'll be fine. It might affect the lower tier, if you like, of maybe composers who don't put the effort in so much, who are just trying to do it really quickly. But then others are thinking it's going to get to the point in maybe a few years' time where it's going to eat away a lot of the opportunities for library music composers and labels.

23:35Tarbuck says that, for the moment at least, TV producers are resistant to AI-generated work. But he's worried that could change. I think most creators in terms of TV and high-end productions don't want anything to do with AI music. And pretty much everyone I know of has said that they're not going to use AI music in their library music. firstly they won't want to be dealing with the legal and copyright issues that arise from using ai music because it's very very murky secondly as creatives they don't really want to support that whole thing anyway and thirdly they want something which is going to align with their brand and their values in the most part they wouldn't think ai music would fit that but as as it gets better as this ai gets better the stuff i played you earlier right now sounds unmistakably human but in a month's time maybe it will be indistinguishable i don't know how

24:25Ludovic Hunter-Tilney: at the moment would you be able to tell the difference between a piece of music made by actual musicians and one that had been produced by a generative AI program? There's kind of an uncanny valley at the moment where you sort of, it might take 30 seconds and then you'll hang on a minute, something about this doesn't sit quite right, the voice sounds a bit weird, but it is getting better and better and better. And I don't want to get complacent about it. You know, we make some really authentic things here with, like say, creaks, mistakes, funny little bits. You might hear someone sort of cheering in the background because they're having a good time.

24:53But all of that stuff can feed into the AI as well. So I don't think anything's safe. And maybe it will be indistinguishable. I really don't know.

25:00Ludovic Hunter-Tilney: Clearly, we are not in the world of gimmicks and toys here. Generative AI is very powerful. It can already mimic man-made music and will presumably get more sophisticated. The great unknown is how much it will compete with human musicians. I guess the answer will partly depend upon how willing we are to listen to AI music. Professional musicians may like to believe that music made by humans will always win out, but they also complain that this is a battle they shouldn't have to fight. None of this would be possible if it wasn't for the fact that there's been a big theft, basically. AI models have been trained on a mass of music that might be copyrighted and used without the creators being asked or even knowing about it.

25:43AI music would not be able to sound like it does today if it hadn't scraped the world's catalogue of music it's a moral travesty really that that's been allowed to happen not to say that ai itself is bad i think like with any technological development it's a double-edged sword people said similar things about synths you know they might take away jobs from musicians they even said about the piano you know hundreds of years ago but this is different because it is taking people's copyrighted intellectual property which should be protected by law and unfortunately that's being completely ignored in favor of developing this technology which is competing with composers.

26:16So their work is being used to compete with themselves. And that just doesn't seem right to me.

26:23Ludovic Hunter-Tilney: This is an argument that has prompted a legal dispute that stands to shape the entire future of generative AI in music. You know, if you want to train on people's music to build these AI models that are going to compete with them, you ask their permission, you pay them, you know, you don't just steal their stuff. Ed Newton-Rex again. They've taken your work without permission. They've trained these models that can essentially do your creative work, and they're offering them pretty much for free. And I think it's made a lot of people very, very angry, right? I mean, they're theft machines. UDO and Suno, the two AI generators mentioned in this episode, are right now facing legal action in the US over their alleged use of copyrighted material to train their models.

27:09Ludovic Hunter-Tilney: The artists and major record labels suing them say this is, quote, copyright infringement on a massive scale. We contacted both AI companies. Suno didn't reply, and UDO said they couldn't comment because the legal cases are ongoing. But in court filings, they've argued that using copyrighted music to train AI models doesn't infringe on copyright. It's fair use. More on that coming up in part two of this series of Tectonic. Are AI music generators really theft machines?

27:52You've been listening to Tectonic, presented by me, Ludovic Hunter-Tilney. The producers are Lulu Smith and Josh Gabert-Doyon. The senior producer is Edwin Lane.

28:04Ludovic Hunter-Tilney: The executive producer is Flo Phillips. Sound design by Breen Turner and Samantha Giovinco. Original, not AI music, by Metaphor Music. Manuela Saragossa is the FT's acting co-head of audio.

29:01Thank you. Conditions change. Alternatives matter. Listen to Strategic Alternatives, available wherever you get your podcasts. Looking for more investing options? Meet SIBO, the exchange that pioneered options trading. With exclusive trading products like VIX and SPX options, SIBO can help you trade in any market environment. There are risks associated with SIBO company products. Review the disclosures and disclaimers at SIBO.com slash US underscore disclaimers.