Technology, Connected - Remove The Beatles From The AI Training Data

Episode Date: November 4, 2025

You're listening to AI-generated music and don't realize it. The musicians whose work trained those models know. They check their empty bank accounts daily.99,000 new songs upload to streaming platforms every day. One in five are AI-generated (Deezer). You wouldn't play a single one at your funeral.59% of musicians use AI in some aspect of their music (Ditto Music). The question: How do real musicians get paid when AI uses their work?We break down Water & Music's research on AI music attribution. Cherie Hu, Yung Spielburg, and Alexander Flores investigated what's at stake and which companies are trying to solve it.The problems:- Session musicians, producers, songwriters—how do they get paid when AI uses their beats?- Record labels can't track which training data influenced which outputs- Proving a musician's input in a model's output is nearly impossible- Copyright law wasn't built for this- Most AI music companies scraped without permission or paymentWe talk about:- How attribution technology could work (vector matching, contribution tracking)- Which companies are building payment systems- Why Suno and Udio's approach creates legal chaos- How ethical companies like Overtune license training data and split royalties- Whether streaming platforms can detect AI-generated music- What happens when the Beatles' catalog trains a modelThe stakes: If attribution fails, AI becomes theft at industrial scale. Music becomes disposable content optimized for algorithms, not humans.Based on Water & Music's research—the best reporting on AI music economics.Share with a music lover.---Research: Water & Music (Cherie Hu, Yung Spielburg, Alexander Flores)Topics: AI music, copyright, attribution, streaming, royalties, training data–(00:00) The Intersection of Music and AI(03:26) Understanding Music Attribution(03:51) Sonic Characteristics and AI Influence(06:39) The Complexity of AI Music Generation(07:36) The Value Equation in AI Music Creation(08:08) Understanding Influence Functions in Music AI(09:44) Challenges of Attribution in AI-Generated Music(11:38) Exploring Embeddings and Their Role in Music AI(14:17) Watermarking and Its Limitations in Music Attribution(15:30) Synthetic Data and Its Implications for Music AI(17:48) Innovative Solutions for Music Rights Attribution(18:01) Distinguishing Compositional vs. Recording Contributions(19:59) The Impact of AI on the Music Industry's Inequities(23:03) Trust and Technology in Music AttributionOther ways to connect with us:⁠Listen to every podcast⁠Follow us on ⁠Instagram⁠Follow us on ⁠X⁠Follow Mark on ⁠LinkedIn⁠Follow Jeremy on ⁠LinkedIn⁠Read our ⁠Substack⁠Email: [e-mail]

It's Ructers and Curious Minds, CEOs, founders. It's Friday. We're live. Welcome to Thinking on Paper. Pocket Edition. In 2004, there were about 99,000 new songs per day uploaded to streaming services. Beezer, which is a French streaming service, a bit like Spotify, they said, AI generated tracks account for 18% of all new songs.

According to a survey by Ditto music, 59% of musicians are used. using AI in some aspect of their music. There's a lot of AI generated music out there. The scary thing for me is that you might be listening to AI music and not even know it. It's not about proving whether AI copied your song. It probably did. It's about identifying whose music it borrowed from, which musicians, which right holders, which publishers, and then recognizing them and making sure that they get credited or paid.

We're reading an article by a company called Water and Music. And the article called How Music, AI Attribution Actually Works. Jeremy is AI attribution, which is what they call it, particularly difficult to music? Is it more complicated for music than cinema or writing? Music in and of itself is really complicated. It's nuanced. And the paper goes through this really in an interesting way. the composition attribution, the recording attribution, right? So then you split into this camp that's really well known in the music industry. You have the publishing side and you have the master side. You know, who wrote the song, who generated the melody and who captured it in a particular instance. And there's influence on that. The producer is an influence on how that's captured. The engineer is an influence on how the settings of the mic pre's and where the mic was placed. And there's a lot of different, a lot of different dependencies that,

that create the output that we, you know, hear on a record or are here on Spotify. You haven't answered my question. So, but you kind of have it, it is more complex. Music has it. Melody, harmony,

rhythm, sound design, you've got the mixing, you got the mastering. When you write a song, could you ever say, okay, I'm borrowing the melody more than I'm borrowing the rhythm.

I'm not borrowing the rhythm, but I'm borrowing this or like, and you don't even know, do you? Because it's all underneath. So you don't know where you're borrowing it from. That's the tricky part, right?

So you, as humans, we listen to music and we, you know, listen to different records and we practice different songs and we learn different songs. And then we start figuring out that the scale from this song, you know, would be really interesting over this particular progression. And there's all this like co-mengling of creative potential. A lot of what these models do is predict, you know, the potential of the next thing happening, right? Yeah, I mean, it doesn't learn from the melody, does it? It learns from the mathematical representation. of that melody. That's kind of like the antithesis of music in a way.

A lot of things would be solved in this grand challenge if we just understood how things were trained. Is it a complete black box? Or is it a model where someone has actually gained permission to use these five songs and gotten permission from the owners of the songs? And then you're able to see how the training data was put together, right? The music industry is going to have to change.

There is no complete solution to this, but it won't be defined by one. product, maybe a collection of these things. And the way we think about rights and who gets paid what is going to have to change somewhat. So let's talk about sonic characteristics and composition and recording and time and data augmentation. Where was your head going when you read this section? Composition. So AI models learn fundamental musical structures like chord progressions, melodic patterns and longer song forms. A model might learn something distinctive like a 146 progression. I prefer 1.45. Without attributing.

it to a specific creator. This raises a critical question. When an AI recreates a common chord sequence who, if anyone, deserves attribution. Okay, I say nobody. I think that you forget chord progressions. You can't, you can't give anybody who's alive any kind of ownership of a chord progression. Can you? Because I read that. Talk to Tom Petty. We can't talk to Tom Petty anymore, but talk to Tom Petty and Sam Smith's. Well, I did say who's alive. Give me a chord progression, which belongs to someone. Well, if you get, that's slippery slope, right? Because then if you start saying chord progressions are, you know,

Pat, which is why they haven't ever been, you start thinking about who. Well, they say it's a critical question. I don't think it's a critical question. You don't think it's a critical question? No, I don't think you can attribute ownership of a chord progression to anybody. No, I agree. I completely agree. But if you do, then you start having to, you know, do the same thing with triads and notes and collections of notes and little melody sequences and then, and chords.

You always say nobody owns a G chord, but I think that Hendrix owns an E minor seventh chord. I think we should give it to him and then give all the proceeds to the Hendrix's estate. I'll tell you what. There's a lot of right in what you just said right there. So recording, like I mentioned before, recording techniques.

She references lo-fi hip-hop. So the characteristics of that recording, you have the vinyl crackle, you have basically all the high-end stuff kind of taken out. Again, how would you attribute that to any one person. This is more about qualification than attribution, I think. So this whole section to me reads as if, all right, here's a song. How do you break up the song into pieces? And then how do you tie those pieces to influence from other songs? You got to chop it up and categorize it.

Question for you on that section. And what's the difference between sampling and what we're talking about here? I agree with what the report says. So sampling, you take a little piece of a section of music and you put it in and use that as a new ingredient. Is it a kick drum? So you sample it and create, you know, put that in for the sound of your kick drum. Or maybe it's a break. You're the A-man break. Funky drummer, the whole thing, right?

But the way AI works is there's so many interdependencies. It's almost like this universe of connectivity between all the pieces and parts that are influencing each other instead of this like one block that goes in to create. It's a Lego block that someone else owns and you put it in as a new ingredient for your Lego block. But it's not interconnected. to all your other ingredients, I think is where I go. I like answer.

The data augmentation thing is really interesting to think about is, you know, the models being trained on particular songs, but then the models generate derivatives of the song by slightly detuning it, maybe speeding up the BPM, adding noise, and then using 100 versions of a song that's manipulated in different ways to train the model. You know, if a really great prompt can generate an awesome result, is that deserve attributes, in the creation of AI music. And the original prompt engineers are, you know, conductors and producers, right, in the studio, generating prompts to get musicians to do something that they really, you know,

pushing the boundaries of what they want to do. I'm not saying a prompt to AI is like Rick Rubin in the studio with musicians. I was just thinking of Rick Rubin. That sounds like a Rick Rubin prompt. It's an interesting thing to think about, you know, whether prompt engineering can be considered creative direction. These are different ways that we could attribute creative inspiration to a AI generated song. Well, not just inspiration, but how it comes together.

The actual production, the composition, all the things that happen in the black box. Yeah, and then the stakes are like who gets the recognition and who gets the reward. This is a way to choose or to create a system where somebody, publisher, writer, songwriter, session, musician, label is acknowledged and rewarding. three different ways do they say influence functions embedding and watermarking I'll influence functions the statistical idea I've got a question for you here so this one says quote a fundamental concept in statistics influence functions measure how much a single data point affects the models output by comparing results with without that point included the principle is straightforward the more a prediction

changes when a specific data point is removed the more influential that point is is considered to be. Okay? And a practical example, picture a music AI that generates a trap beat, influence functions would theoretically measure how different that beat would be if Drake's God's plan were removed from the training data versus removing Kendrick Ramar's Humble. Whichever removal causes a greater change would be considered more influential to the output. Removing outliers from the training data. I mean, what happens to the output if you remove all the Beatles' tracks from the training data. What happens to the app

if you remove all of NWA songs from the training data? It's a very interesting thought experiment or not a thought experiment. I mean, it can be done. You know, you remove all Bob Marley from the reggae data set. What happens in that?

The one drop comes out of everything. Basically, it would, would that check, that would be very interesting because then would it show that we can actually attribute most of the pie to the, to these groundbreaking musicians from the past? Or would something else happen?

How do you think about that bit? Here's what I think about all of this stuff is like, it's like we now have a problem. We have a challenge in front of us, you know, as a music community to figure out how to clean up the mess that was started, right? And the mess is that these companies and these models basically scraped the internet, yanked songs and stole. Yeah, let's just be, yeah,

I keep saying scraped just because it's common language in these publications, but it's really stolen. Like just yanked and stole stuff to make something to try to generate value and sell. And now we're left with the bag going, oh, my God, we have to figure out how to do this. And this idea of leave one out like you're talking about. So we have let's do it very simply, right? So you have a, you have a Jimmy Hendrick song. You have a Bob Marley song. You have a, you know, a Nas.

song and you have maybe a Hank Williams song, right? We push, push a button, generate song, wherever it's prompted you or generate song. There's the output. Okay, we did that once. And now we're going to pull out Telitubbies. See the output again and then judge the influence of the removal of that on the song that this created. But there's so many different songs that can be created using prompts and tracking all of that stuff. And guess what it sounds like to me. You have to literally retrain. the model every time you want to prove attribution, right? So the scalability of this just from a compute perspective is like bonkers, right? Because training takes so many resources right now. So

even if someone wanted to do it, I don't think it would be able to really be done. So they have this thing called approximations that we can approximate the influence. Yeah, I mean, maybe that's some subjective stuff that's a little dicey. But again, this is, I love that they pointed this out because it's an actual tangible statistical thing that we could look at, but definitely has some challenges, right? Yep. Number two, embeddings. So embed whether we're talking about Google search result, a photo app, grouping images of your face or Spotify. Song recommendations, a critical component of modern AI, is the ability to mathematically quantify similarity. For example, a song's raw audio waveform, typically chaotic and computationally unwieldy, might be compressed via embedding into a vector, a string of hundreds or even, thousands of numbers, these vectors act as GPS coordinates in an abstract space. Dude, I've loved this section. Can you explain it to us?

Oh my God, I loved it. Like the ability, this is maybe a quote or close similar, the ability to mathematically quantify similarity, just those words together. That was really, really pretty interesting. And I'm not a math guy. But this is really interesting. The idea you could like suck in text, images and audio. and translate that into coordinates in an abstract space.

But having these vectors as GPS in the space, and the big thing that popped out too is it's more than just longitude, latitude, right? There are hundreds of parameters. But proximity is similarity, right? So did you pull up evernoise.com? I didn't pull up ever noise. Oh, dude.

It was crazy. It was like this. And basically what I, what I, I mean, I looked at it for like five minutes, but it's this mapping of musical genres and how close certain ones are. I didn't even know the names of some of these musical genres that I saw in there. I mean, there are hundreds of them. No surprise, judging by our musical references so far, Jeremy. Fair enough.

Fair enough. But Evernoy's is a great way to kind of think about how this is visualized. But my question for you, are these coordinates the heart of the creative aspect of the song? Is it the mathematical representation of the creativity of these elements that we talked about? No. No. Well, they say themselves, for rights holders, this distinction is critical. Embeddings can flag potential influences what sounds alike, but cannot interrogate the creative process, not why it sounds alike. I think it's a good way and it's a practical way that they could start to attribute ownership, creative inspiration, whatever now. It's not perfect. It doesn't include the why. So it doesn't include, although people can't explain the why, can they? Because it's like Rick Rubin says, it just comes from the universe. from the ether. So it could work as an initial one.

I think it was my favorite in terms of getting something done now. I thought it was really cool. I want to definitely investigate that more in the water and music team want to come on and riff on it. I think that would be a cool subject to talk about. I think better than watermarking. Isn't it? Watermarking. We've had this with AI imagery.

Watermarking, good and good, but just too easy to get around. Also, it can be stripped out. You can literally go into like an EQ filter, find. that audio watermark and and yank it out. There's a lot of, there's a lot of challenges with, with that. But again, it's like, remember, like, I'm not bashing any of these things. We're, we're, we're now trying to fix a messy thing that already happened. And these are kind of really interesting, tangible opportunities to look into that, I think. Yeah, a question for you and the listeners, in fact, if anyone's listening. So,

what do you mean if anyone's listening? They're all listening. You have your radio on and you have that, you use Shazam because, oh, I like that song. I'm going to Shazam it and you get who it is or whatever. Could you add a watermark that says this song is created by AI? When your radio picks that up, it just changes the station. That's, that's, so let me tag on that real quick. So I've talked about this a lot, like an automatic deconstruct mechanism. When in a file of some sort, whether it's video or audio file and say we create a track,

there's a technology that we put into it. Maybe someone can help us develop it. But when someone yanks it and steals it and uses it for another purpose, it automatically scrambles and turns in. to static. I think that would be awesome. That would be awesome. Blockchain, probably. Challenges and the details, I think we've spoken about all of those in our own unique way, except one, synthetic data as companies increasingly use AI generated data to train other AI models.

Attribution becomes a chain of custody problem. I go back to the original stat, 18% of new music on DISA is AI generated. Yeah, that just sounds like a very dark, depressing. synthetic data hole I don't want to really explore too far. Well, we'll just sprinkle on it, right? So you put a song in there, the model gets trained, and then the model takes that and creates 100 versions of your song by changing pitch and doing all kinds of stuff.

And then there are songs created using your song within that model and then variations of those songs. And it's just this cascading like turtles all the way down kind of situation. But yeah, they talk about the challenges, right? The complexity of training these models, is these black box models that we can't see what's going on. Bad tracks, like terribly mixed tracks being able to influence the output of a song.

And actually, they're saying that bad tracks can influence and help models create better music by knowing what not to do. So of those tracks that are submitted, like how many of them are dumpy? And should the dumpy tracks get money? Because they made the model better. Like, it's bonkers, dude.

If people are listening to it, then probably. This is for you. future guests of thinking on paper, who are the key players? While the technical challenges of attribution were made substantial, several pioneering companies are developing commercial solutions that balance future ideals with current market realities. Surreal is the first company. Just, I mean, just super high level, but surreal seems really interesting.

Anytime you're trying to fix a technological challenge, it has to be rooted in what's happening today, and they're rooting in this concept of publishing and recording masters, right in this, and you know, this, that's how the system works today, right? So allegedly they're trying to create technology that separates those and quantifies those influences in a very unique way. Their key innovation is addressing what many consider the central problem in music rights distinguishing between compositional and recording contributions.

Could you unpack that a little bit? The compositional side is literally the lyrics, melody, the chords, that sort of thing, right? And the recording side of it is, you know, that song captured in a moment in time with a producer, with an engineer, with certain musicians. Maybe there's maybe there's musicians that come in that are outside of the entity that wrote it. So you have more influences, you know, the position of the mic, the signal chain, did they take out a chorus? Did they do something different with the bridge? That's the difference in those two things as I think about them. Does that dilute the waters so much that we know.

that Spotify pay pence to most musicians anyway. In that model, does it create too many people to get a slice of the pie and so everybody gets nothing? Well, it's really interesting. And they reference that in this paper. The point system in music between the publishing and the recording master side, they actually had to create a point system that's higher than 100%. It's a 200 point system because there's so many pieces and parts to carve out to that.

but this is going to exponentially affect those pieces and parts and attributions. And if you do it too granularly, you know, you're going to end up getting one cent a year on some of this stuff. So that is the challenge. That's one of the challenges too. Again, you know, the solutions are not the problem here. The problem is stuff getting stolen, used to train models, and now we're trying to fix it, which is a really big endeavor. It's a bigger endeavor than the technology that was created, I think. Music is notorious for the haves and the haves not.

It's a true reflection of society at large. And there's a handful of musicians on Spotify who make all the returns and all of the gigging musicians don't make anything. Maybe it's a question for a later date. But does all of this exacerbate the gap between the haves and the haves not? Is that the long-term effect of AI generated music is that it widens the gap? Well, I'm going to give you a one sentence answer. We talk about it all the time. Technology is jet fuel on whatever it is.

Right. So if it's good or it's bad, it's still jet fuel and it's going to, it's going to accelerate whatever, whatever is his cooking. Musical AI. This is that one of the founders is the former CEO of Beatport. So we're going to connect the dots here. Dean, I think Dean Wilson, who's been on the show, was part of part of that Beatport community, at least as an advisor. But what musical AI is trying to do, it looks like is the idea of creating these royalty sheets with, with influence. And man, what a daunting task that is. Like the output sounds great, but what are the inner workings? How can we, it goes back to trust. Can we trust the inner workings of this? And where they're really, actually, you know what?

We could probably trust this more than most because they're focusing on one stops, which basically are artists that own the publishing and own the masters. So they can make any and all decisions. It's an easier way to get permission quickly instead of having to go to 100 people, you go to one person. We got to. Yeah, we got to land. the plane here, Mark. So Soundverse, they're integrating this, this attribution tech into their platform.

Their platform does a bunch of stuff, right? It can separate stems. It can do text to audio. You can create full tracks. They have this DNA that you could create. Right. So you could create your own DNA mark by uploading your songs. You're giving them permission to train the model with your songs. But your DNA would be the attribution would tie back to this DNA that you created by training the model on. Mark's songs. Yeah, it's like a content part in the program, isn't it? You sign up and then you get royalty payouts based on the data used. There is somebody from Spotify involved in that, so it could be a goer. There's a few others, Lemonade, Lander. So Lander started out as this online mastering tool.

Mastering was, I mean, it's a tremendous art. I remember I grew up with the daughters, the daughters of Rodney Mills. Rodney Mills was one of the top mastering engineers for rock, you know, everything from like 38 special to like fuel hemorrhage in my hand. I think I was sitting in his mastering suite while he was, while he was mastering. It's such a nuanced approach to music to basically it's the last step, right? So you record, you edit, you mix, and then you master.

It's this fine tweaking of levels to make it feel like it hits the right spot. but these guys created an online mastering tool that you could upload and kind of auto master your stuff. I know Dolby has a lot of interesting things related to that as well. But this is an opt-in kind of thing, right? So if you use their tool or you distribute on the Lander platform, it's kind of this 20% of all revenue pool. It's kind of like what Fortnite, what Epic did with Fortnite Islands and you have this creator pool. But it's not really tangible. It's difficult.

like who gets what and why seems a little murky, but would love to talk to the lander folks as well. What does all this mean, Jeremy, for the music industry? I got one sentence for you. Make it a good one. Trust first, then technology. What does that mean for the music industry? So everything is, no matter what technology you create,

if no one trusts it, if no one trusts the flow of the inputs and the outputs and there's no transparency in how the technology is used, that middle black box is not easy to kind of peek inside and say, okay, you know, here's how, here's how attribution flow works. Here's what happens. And it's tricky.

This is super difficult, man. It's a difficult mission, but it's got to be rooted in trust for adoption, for scale, for all of that stuff. But it's a massive challenge. The industry is rooted in a particular way of doing things. The evolution of that is going to be pretty slow, probably from independence outside in. You know, maybe little pockets of some majors considering some of this stuff, but the trust piece. has to be there and transparency for people to get involved with it. And because the precise attribution is impossible, the labels and the musicians and the

platforms can only work with what's measurable. So it's not a perfect fix. There won't be a perfect fix, but the technology that's available, is probably going to lead to lower, royalties for musicians in the long run. Geez, I hope not. I can go in my little studio over here and pull up one of my songs that I wrote and play it and spend hours thinking about what influenced me to create this combination of chords and lyrics and melody and drum sounds and whatever it is. It might be an interesting exercise, but I don't even think I could come up with that because I was done in a moment in time.

Like I can maybe say I like these guys. I have these strategies because... No, no, impossible, because some of that comes from when you were 15 and you were, your parents were driving you to school and a song came on the radio. And at that point, however many years ago when you were 15, something just went in and it resurfaced 35 years later. And you're never going to know, that's the beauty of creation. Maybe on some grand level, like, oh, I like this band. I like this singer.

I like this musician. but not the real creative spark, which I think is what this is all about and how we can't really measure that kind of spark, so we have to build a system of technology that does the best that it can. Here we go as humans trying to wrangle the magic of natural stuff. Holy cow.

I think a unified rights holders win. Does this mean the big labels win? If you own everything, your own rights, then you're going to get a bigger slice of the pie. You're the musician, Jeremy. I don't know. Yeah, I've seen you play some music, Mark. You defer to me a lot. But, well, let's land the plane. Guys, thanks for listening. This has been a pocket edition of thinking on paper. We do a lot of other things. We explore new energy. We explore quantum computing. We talk about AI. We talk about a lot of things. We try to connect the dots to generate some meaning. Mark and I do that individually. We hopefully are helping you guys do the same. So be curious. Stay disruptive. Keep thinking on paper. And I will just add that I don't think that I don't anything defines the human condition better than music and I'm very happy that there are people

out there who are trying to sort out a way to make sure that the people who wrote the songs, the people who made the songs are recognized and rewarded for their time and their effort and their creativity. Totally agree. Totally agree. Good chat. Thanks, water and music. Keep writing your awesome stuff. Come think on paper with us soon. Wow. Wow. That was an emotional episode of Thinking on Paper. If you're still with us, and of course you are after that, if you have any questions about that show, email us at Hello at Thinkingon Paper.xyZ. Subscribe on the platform you're listening and please, as a curious mind, share this with just one person in your life who wants needs would like to hear these conversations. Maybe two. Share it with two. Maybe two. Share it with two.