What Is the Future of Music in the AI Era?

Water & Music asks who, if anyone, deserves attribution when an AI recreates a common chord sequence. No, I don't think you can attribute ownership of a chord progression to anybody anymore.

Clips and outtakes on Instagram ›

In this episode · 18 chapters

  1. It is not about proving AI copied your song
  2. What we are going to cover
  3. What is attribution?
  4. What listening actually means
  5. Examining an AI generated song
  6. Can you own a chord progression?
  7. Sampling
  8. The connection you build with an artist
  9. The little changes every eight bars
  10. Influence, and the stakes
  11. Leave one out
  12. Embeddings
  13. The map of music similarity
  14. The central problem
  15. NFTs, and Tune Welders
  16. Taylor Swift and her masters
  17. What this means for the industry
  18. Trust first, then technology
You go to an art gallery and you look at a Da Vinci or you look at a Picasso and then you look at a copy as a poster of that and your heart doesn’t beat in the same way.
Mark Fielding

We read Water & Music's How Music AI Attribution Actually Works, with Jeremy drawing on his background in music production. We cover publishing against masters, why AI output differs from sampling, and leave-one-out tests that would need a model retrained each time.

Then embeddings that map songs by similarity, watermarks that can be stripped out, and firms trying to pay artists for influence. Mark closes: “And I will just add that I don't think anything for me defines the human condition better than music.”

How Music AI Attribution Actually Works, an article by Cherie Hu, Alexander Flores and Yung Spielburg for Water & Music, sets out how credit could be traced from an AI-generated song back to the music it learned from. Mark Fielding and Jeremy Gilbertson read it on the Thinking On Paper podcast, covering influence functions, embeddings, watermarking and companies such as Sureel and LANDR: “Trust first, then technology.”

What is music AI attribution?

Music AI attribution means identifying whose music an AI model borrowed from and making sure they are credited or paid. Jeremy says the article splits credit the way the music industry already does: the publishing side (who wrote the song) and the master side (who recorded it).

How is AI training different from sampling?

AI training is not the same as sampling, Jeremy argues. A sample is like picking one Lego block out of a bucket; an AI model is like you “melt the Lego blocks into a cauldron”, so every influence is mixed into every other.

What are influence functions in music AI?

Influence functions test how much one song shapes a model's output by comparing results with and without it. The article's example asks how a generated trap beat would change if Drake's God's Plan were removed from the training data rather than Kendrick Lamar's Humble. Jeremy says doing this properly means you would “retrain the model every time you want to prove attribution”, which is why the article turns to approximations.

How do embeddings and watermarks trace AI music?

Embeddings turn a song into a vector, a long string of numbers that acts like “GPS coordinates in an abstract space”, so similar songs sit close together. The article warns that embeddings show what sounds alike but not why. Watermarks, Jeremy says, confirm a contribution but not how much it influenced the output, and can be stripped out with an EQ filter.

Which companies are building music AI attribution tools?

The article names companies building attribution tools, including Sureel, Musical AI, Soundverse and LANDR. Jeremy says Musical AI works with “one stops”, artists who own both their publishing and masters; Soundverse lets musicians train its model on their own songs to create a personal “DNA”; and LANDR, which began as an online mastering tool, runs an opt-in pool he describes as 20% of all revenue.

Where Mark and the article disagree

The article calls attribution for a common chord sequence a critical question. Mark disagrees: “I don't think you can attribute ownership of a chord progression to anybody anymore.” Jeremy agrees, warning that once progressions are owned you start doing the same with triads and single notes.

Questions

Who wrote How Music AI Attribution Actually Works?

How Music AI Attribution Actually Works, an article by Cherie Hu, Alexander Flores and Yung Spielburg for Water & Music, sets out how credit could be traced from an AI-generated song back to the music it learned from.

How does LANDR pay artists for AI training?

Jeremy says Musical AI works with “one stops”, artists who own both their publishing and masters; Soundverse lets musicians train its model on their own songs to create a personal “DNA”; and LANDR, which began as an online mastering tool, runs an opt-in pool he describes as 20% of all revenue.

What does Mark say defines the human condition?

Mark closes: “And I will just add that I don't think anything for me defines the human condition better than music.”

Transcript

Mark Fielding disruptors and curious minds, CEOs, founders. It's Friday. We're live. Welcome to Thinking on Paper, Pocket Edition. No guest on a Friday. It's just me and Jeremy speaking about the latest in a technology. And today we're speaking about something close to both our hearts. We're talking about music and AI created music and the attribution of ownership. When you listen to an AI song,

How do you know who is the, where's the creative spark come from? Who's getting their award? Where's the...

Who shaped it? Who did they copy from? Where's the creative inspiration? And there are ways to make that a little bit clearer, to make it a little bit easier. We're reading an article by a company called Water and Music and the article called How Music AI Attribution Actually Works. Jeremy, since your background is in music production, I'm going to pass it over to you.

Jeremy Gilbertson Yeah, I think there are a couple of good clarifications that we want to make on this one too. Well, first of all, Water & Music, I've been following them for a long, long time. think as soon as Web3 started kind of spinning its head around years ago, they were ahead of a lot of things that I was really interested in. They do such a great job of organizing very technical information and turning it into something that is very tangible to someone who's outside.

a particular ecosystem. Would love to have that team on the show at some point. I think it'd be a fun episode. ⁓ a of call outs on the authors then because ⁓ Sherry, Hugh, Alexander Flores and Young Spielberg are the three authors of this essay. So their work and it's very in depth. There's a lot of research going on. Links below to read it for yourself. But yeah, they've done a lot of work on this. So shout out.

think on paper with us. ⁓ Okay, so the clarification that I wanted to be sure we got out in the open really quick because music and AI is a ⁓ very touchy subject. ⁓ And our, or at least my belief, I'll let you speak to it, Mark, but my belief is when there's general permission for use to train models, to do things like you would do ⁓ in a film, in a video game. You wouldn't just yank someone's music and stuff it in there and call it your own. You would go to that person and say, hey, can I use this for this particular purpose? So this isn't meant to be a conversation of should AI music even exist? Because guess what? Sorry, guys. It's out of the bag. And just like with a lot of things with AI, we talk about with the Empire of AI and Book Club that we wrote.

that we read, Karen Howe wrote it. It's a brilliant book. We kind of broke it down, but the cat's out of the bag with this. It's out of Pandora's box. People are attempting to wrangle it in a lot of different ways, not only usage, but ⁓ policing this thing too. So the purpose of this discussion, I think, is rooted in how do you credit the people, even though sometimes things were done without permission, how do you figure out how these little micro, she calls it micro influences ⁓ to generate these AI tracks. So I wanted to start with that.

It is not about proving AI copied your song

Mark Fielding Yeah, 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. And there's obviously a lot of caveats to that. How do you decide who the importance of the influence if you have a track, how do you decide which musicians played the biggest role in that? And this goes back to the beginnings of music because I mean, if you listen to an AI created track made by a new rock band or an AI rock band that's heavily borrowed from the last 30 years of rock music and those bands and those musicians from the past 30 years of rock borrowed from what came in the 50s and the 60s. They borrowed from Paul McCartney. Paul McCartney borrowed from Beethoven. So if you're you're creative, AI is creating a song today that the origins are from 300 years ago and it's traveled through all those generations of music. How do you recognize the most important parts of that? And how do you make sure that those people are credited and rewarded? And yet, as we learn, it's very difficult. Question for you, Jeremy. Is AI attribution, which is what they call it, particularly difficult to music. Is it more complicated for music than cinema or writing?

Jeremy Gilbertson Yeah, I mean, music in and of itself is really complicated, right? It's nuanced. ⁓ Sure, have a progression. 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, right? 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 dependencies that create the output that we hear on a record or hear on Spotify. ⁓ The complexity of it as it relates to other things and other...

It's difficult. can just say it's complex, right? It's super complex, it's super nuanced. There's a lot of different pieces and parts and they do a great job of outlining some of this stuff. I haven't really answered, yeah.

Mark Fielding you haven't answered my question. So you kind of have it. It is more complex, certainly than a book or a novel. Maybe cinema has a lot of different inputs, so maybe that's a little similar vein. But music has a melody, harmony, rhythm, sound design. You've got the mixing, you've got the mastering, you've got like where...

You write a song, you've written music, when you write a song, which parts are you borrowing from your many years of experience? 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.

Jeremy Gilbertson the tricky part, right? So you, as humans, we listen to music and we 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 would be really interesting over this particular progression and there's all this like co-mingling of...

creative potential, which it does the same thing in these models, all of these different interconnects. And I want to get into that because she illustrated something really cool in the idea of vector in a particular space and how these patterns are actually connected. But if you remember, a lot of what these models do is predict the potential of the next thing happening, whether it's a word, whether it's a vector, whether it's a ⁓ voltage.

Jeremy Gilbertson that affects a spin on a qubit like we talked about this week with quantum computing, right? ⁓ learn from the melody, does it? It learns from the mathematical representation of that melody.

Mark Fielding that, yeah, that's kind of like the antithesis of music in a way.

Jeremy Gilbertson Yeah, but mapped out. there's a couple of interesting ways that some of those vectors are mapped out. But before we dive into that, let's kind of start from the top. this is a pretty big statement that I kind of pulled out is the idea of linking the AI outputs back to specific training inputs that influence them. Now, if.

You know, we've been studying AI for a really long time, not quite since the 1960s, but you know, over the last few years, we've, we've really dove into it, right? ⁓ it's really difficult to figure out how models are trained. Like it's the whole black box thing. So, you know, people aren't really open to showing, Hey, here's how my model is trained, which is one of the problems I think, like Karen Howe's book really presented the idea of like, Hey, a lot of things would be solved.

in this grand challenge if we just understood how things were trained, right? ⁓ So it's this idea of causal relationships between inputs and outputs. This idea of interpretability and explainability I thought was really interesting. ⁓ kind of knew of it a little bit, but this is really succinctly ⁓ explained with models. Explainability is why a model makes a particular choice or decision, right? And then interpretability, of it is the how, right? It's how easy is that model to understand how it operates, right? 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? So those two things are pretty interesting.

What we are going to cover

Mark Fielding Yeah, well, do you mind if I just lay out what we're going to talk about in this article just so we can keep it organized? So part one, how music attribution actually works. What is attribution? That's the first part. Then how do we attribute? What are the different strategies and ways that we could attribute? Then who gets paid and how?

Jeremy Gilbertson You keep us organized, man. Let's stay on that track. Yeah.

Mark Fielding do we attribute, they speak about some of the key players in the game. There's a lot of names, a lot of companies that basically I was reading that thinking that they should be a guest on thinking on paper. They should be a guest on thinking on paper that it really does lay out and give us the names of the companies who are trying to solve for this problem. And then we get to the conclusion and the conclusion is essentially that.

The music industry is going to have to change. There is no complete solution to this. There are different ways and different holistic models to get it, 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 for is going to have to change somewhat. But let's go back to the beginning. What? Yeah, good luck.

Jeremy Gilbertson Yeah, good luck, good luck, good luck. Well, it's unfortunate to say that. Well, before we get into what's attribution, ⁓ there's that one word that popped up at the very end that is such a through line through all of this stuff and it has nothing to do with technology. The widgets, the pieces, the hardware, the algorithms, the cube, whatever you're dealing with, starts with a T, ends with a T, and it's rusty in the middle, right? ⁓

What is attribution?

Mark Fielding What is attribution?

Jeremy Gilbertson Trust, so it's, I mean, that's the big, one of the biggest things, but that's a human challenge, that's not a technology challenge. And technology only exacerbates the presence of it or lack of it, right? So speaking of inputs and outputs, the human input needs to influence this piece of the puzzle, which is gonna be a challenge.

Mark Fielding I've got a little bit of data to give you, which will help us as we speak through this to realize how important what we're speaking about is. These are very rough statistics, bit of data that I picked off the internet. So speaking of trust, But quote, in 2024, there were about 99,000 new songs per day uploaded to streaming services. 99,000 per day.

between 23 and 25, that number varies a little bit, but ballpark figure at 99,000. Deezer, which is a French streaming service, a bit like Spotify, they said AI generated tracks account for 18 % of all new songs on that platform. And according to a survey by Ditto Music, 59 % of musicians are using AI in some aspect of their music.

That's huge. It's only 2025. That's obviously gonna get bigger. There's a lot of AI generated music out there. The scary thing for me is that, although I don't listen to streaming services anymore for this reason and other reasons, but you might be listening to AI music and not even know it.

Jeremy Gilbertson Wait, wait, wait, pause. You don't listen

Mark Fielding I have unsubscribed from Spotify and Deezer and all these things, YouTube, but I have no ads on YouTube, so I have to choose what I'm listening to. Or internet, internet, internet radio.

Jeremy Gilbertson How do you play music?

YouTube could count as a streaming service in a way. Yeah.

Mark Fielding Yeah, it could. Okay. What I mean, one of those algorithm fed. Okay. Well, I don't have any ads, so I can't, it doesn't recommend tracks on YouTube and mostly internet radio.

So you don't have to worry about your kids grabbing your Spotify and deconstructing your algorithm to have silly songs about animals versus... ⁓ about me choosing what I want to listen to rather than being fed. ⁓ Well, increasingly AI generated music of 18 % of the tracks on Deezer, of the new tracks are AI generated. But that's just my opinion. That's just what I do. You got to...

What listening actually means

Jeremy Gilbertson Okay, a couple of callbacks, to kind of set the tone before we dive into some of these sonic characteristics and how AI is deconstructing those patterns and that sort of thing. I wanna also think about when plugins first came out, when Pro Tools first came out years ago and then you had this rapid... ⁓ introduction of these plugins where you could actually, you know, pull up a particular, you know, like a tube tech mic pre instead of the one that's, you know, sitting down here in my rack that I could physically manipulate by myself. These plugins started to model the sound of that particular physical mic pre. And then as those plugins got better, they had presets that's you you didn't really have to understand compression ratios and all of that kind of stuff. You just selected the preset that you wanted to have happen. So I think, I don't know that it's actually entirely applicable to, you know, AI music composition, but there are, there's technology assisted music production that's been happening for, for a really long time. Um, and I don't think, I don't think a lot of those folks, you know, who came up with that really cool, rich compression ratio using a particular technology. When that got converted into a preset, I don't know that anyone got any money for that. So that's something to something to consider and think about. And also I'm to bring up, I bring up a lot of books, Mark and, and this particular book was, was really powerful as we start talking about like the categories to, um, to put music into, right? When you're trying to figure out the pieces and parts of music, you gotta separate it into categories. Vic Wooten wrote an amazing book called The Music Lesson that I reference a lot and I always give it to people who don't understand music that wanna understand music a little bit better. But he has 10 fundamental elements of music that he talks about and it's this brilliant story between, you know, it's like this Mr. Miyagi character that comes to visit him.

and walks him through all of these in like a wax on wax off kind of way. It's really cool. But ⁓ we're talking about notes, we're talking about articulation, technique, emotion, dynamics, rhythm, tone, phrasing, space is one of my favorite ones, and listening. So space is the idea of lack of sound, right? That creates interest. So all of these 10-funt-listening is for a reminder here.

Mark Fielding What does he mean by listening?

Jeremy Gilbertson Oye says the act of truly hearing and understanding the music. So it's a little bit vague, I think understanding it to create it, I think that's a vague one, but phrasing is like a sentence, like a musical sentence, right? Space is how silence can generate interest in music, right? You the notes don't have to be there. Sometimes the lack of notes makes something interesting, right? Rhythm is pattern of notes. Tone is the quality of the note, right? So anyway, long story short, that...

I went to that right when she started talking, or right when, I guess they, because it's a group of authors, started talking about how to quantify the little pieces of music. 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?

Examining an AI generated song

Mark Fielding When examining any AI-generated song, multiple dimensions of musical influence come into play simultaneously, composition, recording and time. 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 1-4-6 progression. I prefer 1-4-5. 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. can't give anybody who's alive any kind of ownership of a chord progression. Can you? Because I read that and I thought that isn't to Tom Petty, we can't talk to Tom Petty anymore, but talk to Tom Petty and Sam Smiths. Right, right, right.

Mark Fielding Well, I did say who's alive.

Can you own a chord progression?

Mark Fielding mean, is anybody... Give me a chord progression which belongs to someone.

Jeremy Gilbertson Well, if you get that slippery slope, right? Cause then if you, if you start saying chord progressions are, you know, Pat, which is why they haven't ever been, you start thinking about who.

Mark Fielding Well, they said it's a critical question. I don't think it's a critical question.

Jeremy Gilbertson You don't think it's a critical question?

Mark Fielding No, I don't think you can attribute ownership of a chord progression to anybody anymore.

Jeremy Gilbertson I completely agree. But if you do, then you start having to do the same thing with triads and notes and collections of notes and little melody sequences and and and chords, you know, right? So. ⁓ 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 estate.

Jeremy Gilbertson I tell you what, there's a lot of, there's a lot of ⁓ right in what you just said right there. ⁓ Okay, so these little pieces, right, you have ⁓ these chord sequences, right, and then you add different keys to that, right, so you have different options. The mathematical options are endless, right, with these combinations. Well, actually, if you think about it, they're not infinite, it's probably finite, but there's a whole bunch of them. ⁓

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, you know.

Mark Fielding how would you attribute that to any one person?

Jeremy Gilbertson No, I think I don't think it's this is more about qualification than 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 from other songs? You got to chop it up and categorize it first. ⁓ Yeah.

Mark Fielding Question for you on that section and Jamie, what's difference between sampling and what we're talking about here? Because sometimes the argument crops up that, but samplers, hip hop's been doing this for years with sampling, but is that the same thing?

Sampling

Jeremy Gilbertson I think, no, I agree with what the report says. So sampling, take a little piece of a section of music and you put it in and use that as a new ingredient, right? So is it a kick drum? So you sample it and create, know, put that in for the sound of your kick drum or maybe it's a break. You know, the amen break, you know, funky drummer, the whole thing, right? ⁓ But the way AI works is there's so many interdependencies kind of commingled that it is a,

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. 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 with it.

Mark Fielding answer. One of my favourite albums of all time, Jeremy, Three Feet High and Rising, Delasol. And you couldn't listen to that album for a long time on any of streaming platforms because they used so many samples that they couldn't attribute those. But every sample they used, and I think they did eventually, identify it and that particular piece of sound was used in a different way.

Mark Fielding We're not talking about that, we like that sampling. is actually kind of taking the ideas and what you have, the output doesn't necessarily rep or mirror the input in any way. So it's something very different to sampling.

Jeremy Gilbertson sampling is, you you throw the Lego blocks in a bucket and you pick a Lego block up at a time. And then, you know, what, what we're talking about is you literally melt the Lego blocks into a cauldron and stir them around and they're all melted together, I think is, is, is the way to think about it. Um, let's hit a couple more of these. Yeah. Yeah.

Mark Fielding Can I just another question for you on that?

Because I was just thinking about three feet high and rising and it's not connected to this. But do you think...

that an AI musical generation tool could ever reproduce something like three feet high and rising, which is quite like an interesting question because three feet high and rising was built on samples and AI is this replication machine. But I put out there that no, couldn't ever produce something with such awesome musical power as that.

The connection you build with an artist

Jeremy Gilbertson Well, think about this. So I keep calling back to the show we recorded this week about quantum, right? So ⁓ Brandon was our guest and ⁓ he is working with spin qubits, right? And he basically talked about working on spin qubits on their own as a researcher versus having an AI system to enable the ability to scale working on many qubits and creating many versions of him to do that.

It's it's tough. think that I think the difference is when I hear a song that I'm connected to there's something there's something about knowing that it's from flesh and blood human comes from here comes from here that I don't think will ever Will ever lose I hope we don't ever lose the ability to connect to an artist in a particular way like when you go

Jeremy Gilbertson When you hear a song and you go see that artist live, you build that connectivity, you build that, I don't know, that connection is that I don't think I'll ever get.

Mark Fielding Yeah. And listeners can experience that anytime they want. You go to an art gallery and you look at a Da Vinci or you look at a Picasso and then you look at a copy as a poster of that and your heart doesn't beat in the same way. There's something, like you said, that connects you to the artist and that connection between you and the artist. Maybe it's more powerful when the artist is

400 years dead, I don't know. But yeah, there's something about the real McCoy, does it matter? that raises the question with this. mean, if 18 % of songs on Deezer are AI generated, if you don't know...

you don't know and it doesn't matter. Is it the knowing which creates the connection?

Jeremy Gilbertson I want to call out another couple of thinking on paper dictionary words that pop up a lot. The first one was trust. The second and third one are awe and wonder. And I think what you're talking about is the activation of awe and wonder by being around the actual Da Vinci, by putting on a Demiola McLaughlin vinyl where you sit down and you go, holy cow, these guys literally just ripped out.

on a stage in San Francisco in the 70s and they're doing stuff on a guitar that I can't even think about doing. Like captured moments of humans doing real shit. There you go. Captured moments of humans doing real shit. Okay. ⁓ Let's quick hit the rest of these. So time is another kind of category or qualifier, right? Because music, this is almost a quote, music unfolds over time. Absolutely like the...

Mark Fielding Yeah. Okay. Let's go back to the thing.

The little changes every eight bars

Jeremy Gilbertson the best songs have these interesting little changes, say every four bars or so that keeps your interest, right? Whether something drops out or something comes in, it's like, here's a new guy to the party. It keeps and holds your interest. The data augmentation thing is really interesting to think about is 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 a hundred versions of a song that's manipulated in different ways to train the model, then man, that just gets like super convoluted and ⁓ challenging to figure out, is like convoluted is the word, isn't it? They add layers and layers. It's not just about the audio files yet. Data augmentation. They talk about prompt engineering, multimodal influences, model design. Yeah, it's convoluted is a good word, I think.

Jeremy Gilbertson Yep. And then, and to finish this section, you know, they, they reference, you know, prompt engineering and, know, could that really be, um, you know, if a really great prompt can generate an awesome result, you know, does that deserve attribution in the creation of AI music and you're the original prompt engineers are, 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. Um, I'm not saying a prompt.

to AI's like Rick Rubin in the studio with musicians, but.

Mark Fielding I was just thinking of Rick Rubin. That sounds like a Rick Rubin prompt.

Jeremy Gilbertson Right, right. But you know, it's an interesting thing to think about, you know, whether prompt engineering can be considered creative direction, right? ⁓ And then they talk about some cross modal stuff, but let's go to this value distribution piece, right? So think about these questions, Mark, you know, what are the music elements that you want to track? You know, let's point back to Vic Wooten, you know, which influences the output enough to generate compensation and how do you track it with some confidence?

are the things that this section talks about. ⁓ How do you feel about this value equation and what would you change?

Influence, and the stakes

Mark Fielding Can I just like clarify that because it's a bit murky. Like, so what, like, how do we attribute? Like, so these are different ways that we could attribute creative inspiration to a AI generated song. Is that correct?

Jeremy Gilbertson Yes. Well, not just inspiration, but how it comes together, right? The actual production, the composition, all the things that happen in the black box.

Mark Fielding Okay, I like some of Yeah.

Jeremy Gilbertson 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. So they...

Mark Fielding three different ways, 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 a model's output by comparing results with and 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 considered to be. Okay. 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 Lamar's Humble. Whichever removal causes a greater change would be considered 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 output 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 you remove all Bob Marley from the reggae data set and what happens in that?

Jeremy Gilbertson The one drop comes out of everything.

Mark Fielding Basically, that would be very interesting because then would it show that we can actually attribute most of the pie to these groundbreaking musicians from the past? Or would something else happen? How do you think about that bit?

Leave one out

Jeremy Gilbertson 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 as a music community to Figure out how to clean up the mess that was started right and the mess is that? These these companies and these models basically scraped the internet yanked songs and

Jeremy Gilbertson stole, yeah, let's just be, yeah, forget, I keep saying scraped just because it's common technology, or 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, my God, we have to figure out how to do this. so this influence function, it's a statistical thing, it's well known.

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 Jimi Hendrix song you have a Bob Marley song you have a you know a NOS song and you have ⁓ Maybe a Hank Williams song, right? So yeah, I know we're we're riffing And maybe it's maybe a Teletubbies song at the very end or something whatever it is, right? You remember that baby Einstein music or whatever though? ⁓

Mark Fielding Can we get something for the kids, Jeremy? Something for the kids. We've got a Taylor Swift song.

Jeremy Gilbertson So you have this and it starts generating, know, we push a button, generate song, whatever it's prompted, generate song, boop, there's the output. Okay, we did that once and now we're gonna pull out Teletubbies, boop, see the output again and then judge the influence of the removal of that on the song that's 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. 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?

Embeddings

Mark Fielding Yep, number two, embeddings. So embed, whether we're talking about Google search results or photo app grouping images of your face or Spotify song recommendations, a critical component of modern AI is the ability to mathematically quantify similarity. So embedding.

For example, a song's raw audio waveform, typically chaotic and computationally unwieldy, might be compressed via embedding into a vector, like a string of hundreds or even thousands of numbers. These vectors act as GPS coordinates in an abstract space.

Jeremy Gilbertson Dude, I loved this section. Oh my God, I loved it. the ability, this is maybe a quote or a 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. you know, I'm not a coder, I'm not an equation guy, but this is really interesting, the idea that you could like suck in

Mark Fielding Can you explain it to us then?

Jeremy Gilbertson text, images, and audio and translate that into coordinates in an abstract space. And when I hear that, our heads are around quantum right now. I'm thinking Hilbert space is kind of the first weird thing that popped up. And maybe there's some similarities between things interacting in a Hilbert space and the vectors here.

Mark Fielding This is Stig on paper, there's always connections, Jeremy. Well done.

Jeremy Gilbertson that's what we're here to do, right? ⁓ 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 that, it's not just an X, Y axis, there's like depth and breadth and all of that stuff. And really, I mean, that's kind of, that's like a musical Hilbert space is kind of where we're landing with this stuff. But proximity,

Mark Fielding It's a Hilbert space.

The map of music similarity

Jeremy Gilbertson is similarity, right? So did you pull up evernoise.com?

Mark Fielding I didn't pull up ever noise.com.

Jeremy Gilbertson it was crazy. It was like this, and basically 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 were hundreds of them, but it had.

Mark Fielding judging by the musical references so far, Jeremy. ⁓ fair enough. But Evernoise 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?

Jeremy Gilbertson 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, which, okay, let's not get bogged down in quantum, but that's maybe where the maths of quantum comes in at a later date. But yeah, it's 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. will. It doesn't include the why, so it doesn't include, although people can't explain the why, can they? Because it's like Vic 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.

Jeremy Gilbertson I thought it was really cool. I want to definitely investigate that more and if Water and Music team want to come on and riff on it, I think that would be a cool subject to talk about. ⁓ than watermarking, isn't it? Watermarking, we've had this with AI imagery, watermarking, good, but just too easy to get around.

Jeremy Gilbertson Yeah, well, just, the quick piece of that is it confirms the contribution, but not the level of influence, right? ⁓ So also it can be stripped out. know, audio, like nefarious people wanting to grab stuff and steal it and that sort of thing, you can literally go into like an EQ filter, find that audio watermark and yank it out. There's a lot of challenges with that. But again, it's like, hit remember like I'm not bashing in any these things. 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.

Mark Fielding A question for you and the listeners, in fact, if anyone's listening. if you listen, if you have your radio on and you have that, you use Shazam because, I like that song. I'm going to Shazam it and you get who it is or whatever. And it tells you because there's watermarks in the signal that's been sent out. Could you add a watermark that says this song is created by AI and when your radio picks that up, it just changes the station. Right.

Jeremy Gilbertson What do mean, if anyone's listening? They're all listening.

That's, 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 into static. I think that would be awesome. Yeah.

Mark Fielding probably. Challenges in the details. I think we've spoken about all of those in our own unique way, except one, synthetic data. 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 Deezer is AI generated. Yeah, that just sounds like a very...

dark depressing synthetic data hole I don't want to really explore too far. Any thoughts on it Jeremy?

Jeremy Gilbertson Well, I'll just, let's just, 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 a hundred 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, kind of situation. ⁓ but yeah, they, they talk about the challenges, right? The complexity of training these models, 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 it actually They're saying that bad tracks can influence and help models create better music by knowing what not to do so of of those a 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

Mark Fielding If people are listening to it, then probably. Okay, this is for you, future guests of thinking on paper, who are the key players? While the technical challenges of attribution remain substantial, several pioneering companies are developing commercial solutions that balance future ideals with current market realities. Surreal is the first company. Did you check out any of these, Jeremy? Did you have time or not?

Jeremy Gilbertson I did, yeah, 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? 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. So I didn't go beyond that, but that would be really interesting to unpack some stuff with those guys.

The central problem

Mark Fielding 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?

Jeremy Gilbertson So if we go back to the section we talked about in like when you take a song and you separate all the pieces and parts to quantify the connections of influence to AI generated music, the compositional side is literally the lyrics, the melody, the chords, that sort of thing, right? And the recording side of it is, 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, the, ⁓ did they take out a chorus? Did they do something different with the bridge? ⁓ So that's the difference in those two things as I think about them.

Mark Fielding We'll get to a conclusion, obviously, a little not pushback. We haven't been there. I've been on the show yet, so I don't know. But does that dilute the waters so much that we know that Spotify pay pence to most musicians anyway? If in that model, does it create too many people to get a slice of the pie? And so everybody gets nothing.

Jeremy Gilbertson Well, it's really interesting. They, and they referenced that in, this, in this paper, this article that, ⁓ the point system in music between, the, the, publishing and the recording master side, they actually had to create a point system that's higher than a hundred percent. It's a 200 point system because there's so many pieces and parts to carve out to that. But this is, this is going to exponentially affect those pieces and parts and attributions. And if you, and if you do it too granularly, you know, you're going to end up getting

0.00001 cent a year on some of this stuff. So that is the challenge. That's one of the challenges too. Again, 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.

Mark Fielding Music is notorious for the haves and the have 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 have not? Is that the long-term effect of AI generated music is that it widens the gap? I don't want you to answer that now. I'm just putting it out there as...

Jeremy Gilbertson 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 accelerate whatever is 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 that Beatport community, at least as an advisor.

NFTs, and Tune Welders

Jeremy Gilbertson And one of the NFTs, and this is, I'm bringing up an NFT ⁓ that was really pretty cool, just the idea and the concept of it. And I actually got one. It was a Synth Head ⁓ NFT, which I thought was just a really cool, creative piece of the pie. Here's me saying something good about Web3, Mark. ⁓ But what musical AI is trying to do, it looks like, is the idea of creating these royalty sheets with influence. And man, what a daunting task that is. Like the output sounds great.

Jeremy Gilbertson but what are the inner workings? How can we, it goes back to trust. Can we trust the inner workings of this? ⁓ Where they're really, actually, you know what? We could probably trust this more than most because they are focusing, this is basically just my quick review, but 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 like this library.

that my company, Tune Welders, which is now shut down, but like we created a music library called Tune Well and we approached artists that own the publishing and own the masters to get permission to use their stuff. It's an easier way to get permission quickly instead of having to go to 100 people, you go to one person. ⁓ So the royalty sheets on that, actually that might be an interesting one to track now that I think more about it.

Mark Fielding ⁓ that's speaking of Dean Wilson, that dead mouse has bought those. I think Taylor Swift has just bought back all her masters and didn't quite a lot of the bigger artists have sold a lot of their

Taylor Swift and her masters

Jeremy Gilbertson Taylor Swift actually bought back her masters and re-recorded those records, which I thought was such a bold and interesting move. ⁓ Yeah, we gotta land the plane here, Mark. So Soundverse, this is kind of interesting, right? ⁓ So they're integrating this attribution tech into their platform, and their platform does a bunch of stuff, right? It can separate stems, it can do text to audio, you can create full tracks.

Jeremy Gilbertson based on a quick review, you know, this is more of a hobby musician kind of thing to play around with and 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.

Mark Fielding Yes, like a content partner program, isn't it? You sign up and then you get royalty payouts based on the data used, don't you? So, although there is somebody from Spotify involved in that, so it could be a goer. There's a few others, Lemonade, Lander. I was actually speaking, I'm speaking to somebody from Lander on LinkedIn. So watch this space, but I think we can get somebody connected to them.

Jeremy Gilbertson So Lander started out as this online mastering tool, right? Mastering was, I mean, it's a tremendous art. I remember ⁓ I grew up with the daughters of Rodney Mills. Rodney Mills was one of the top mastering engineers for rock kind of back in the day. everything from 38 Special to Fuel.

hemorrhage in my hand, think I was sitting in his mastering suite 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 and 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. 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 they, you have this creator pool, but it's not really, 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 this means for the industry

Mark Fielding Okay, what does all this mean, Jeremy, for the music industry at large?

Jeremy Gilbertson I got one sentence for you.

Mark Fielding Make it a good one.

Trust first, then technology

Jeremy Gilbertson Trust first, then technology.

Mark Fielding What does that mean for the music industry?

Jeremy Gilbertson 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, here's how the attribution flow works, here's what happens. And it's tricky, this is super difficult, man. It's a difficult mission, but it's gotta be rooted in trust for adoption, for scale, for...

⁓ All of that stuff, but it's a it's a massive challenge. The industry is 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 for people to get involved with it precise attribution is impossible. The labels and the musicians and the platforms.

Mark Fielding can only work with what's measurable. So it's not a perfect fix. There won't be a perfect fix, but the technology is that's available.

is probably going to to lower royalties for musicians in the long run.

Jeremy Gilbertson Jeez, I hope not.

Mark Fielding Another one that they put, sorry.

Jeremy Gilbertson Well, the tricky thing like, yeah, no, just the tricky thing, just to get your head around it. Like I can go, ⁓ you know, 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. Cause that was done in a moment in time.

Like I can maybe say I like these guys. I have these strategies because.

Mark Fielding no, no, impossible, because some of that comes from when you were 15 and your parents were driving you to school and a song came on the radio. 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.

you're never going to be able to maybe on some grand level like, 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 creative spark. So we have to build a system of technology that does the best that it can. ⁓

Jeremy Gilbertson Here we go as humans trying to wrangle the magic of natural stuff. Holy cow. think that...

Mark Fielding rights holders win does this mean the big labels win if you'd like you own everything you own with rights Then you're gonna get a bigger slice of the pie

You're the musician, Jeremy, I don't know.

Jeremy Gilbertson Yeah, I've seen you play some music, Mark. You defer to me a lot. 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 have a book club. ⁓ Actually, Mark, you want to announce the next book? Is that applicable?

Mark Fielding can't remember the names of the authors. It's called Room for Space.

Jeremy Gilbertson Okay, we'll do that later. We've got a new book that we're focusing on space tech again. We're on a tear with that. ⁓ We explore new energy, we explore quantum computing, we talk about AI, we talk about a lot of things, and we try to connect the dots to generate some meaning. and I do that individually. We hopefully are helping you guys do the same. So be curious, stay disruptive.

Mark Fielding Keep thinking on paper. And I will just add that I don't think anything for me 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.

Jeremy Gilbertson Totally agree, totally agree. Good chat, thanks Water & Music. Keep writing your awesome stuff. Come think on paper with us soon.