Episode 106: Can AI And Music Rights Coexist?

AI music tools are moving fast, but one issue keeps poisoning the well: models trained on copyrighted songs with little or no visibility into what went in. When creators cannot see the training data, they cannot consent, negotiate, or get paid, and we end up stuck in a bleak loop of scrape first and litigate later. I want a better outcome, and I think we already have a blueprint hiding in plain sight.

I unpack how legal sampling works through TrackLib and why it is more than a niche producer tool. It is a rights and provenance system: pick from a cleared catalogue, create, clear before release, and ensure credit and compensation reach the right people. Then I map that logic onto generative AI in the music industry with a licence-first approach: opt-in training catalogues for labels, publishers, artists, and other rights holders; a training provenance database that shows what was included; revenue participation through royalties or usage-based payouts; influence reporting; and clear commercial rights so users know what they can release.

We also face the tough questions head-on: scale, value allocation across millions of training tracks, the difference between influence and copying, and the competitive pressure that makes unlicensed scraping tempting. I point to early proof-of-concept signals, including companies pursuing licensed training and partnerships with major rights holders and organisations, and I explain why moving beyond a flawed declaration model matters.

If you care about music copyright, AI training data transparency, and building fair royalties for creators, listen through and tell me where you land. Subscribe, share with a creator friend, and leave a review so more people can join the conversation.

Transcript

Welcome And The Big Question

SPEAKER_01 0:00

Welcome to the Music Business Body, the podcast decoding how the modern music industry actually works. I'm Johnny Amott, industry consultant, artist manager, producer, educator, and author of the music business for music creators. Every Wednesday I explore the ideas, technology, and industry shifts shaping the future of music. This show helps creators navigate the rapidly changing business of music.

The Training Data Transparency Problem

SPEAKER_01 0:57

But perhaps the biggest problem in the music industry is the idea of big tech companies training their AI models on pre-existing copywritten material and then not telling people what the actual training data is. I'm not going to name them, but you know who they are. Today I am not here to kind of give a big backlash on that. Instead, I am here to present to you what I think could be actually a remunerative solution to this going forward. So let me get into it. If we think about AI as a multi-dimensional, multi-faceted version of sampling, right? So when we sample something, we might take a piece of it and we go, right, we've taken that little piece of your work over there. So therefore, we're going to credit you for that, whether that's on the song or the recording or both, and then the rest of it we keep, right? That model there could work in AI. And I'm going to present how I think that could actually happen. So what we're going to do is to start off by looking at an existing model that we can actually learn from that already sort of does this, but in sampling it. Then we're going to look at how that can actually transfer itself over to AI models. Stick with me on this, everybody. This could be a game changer.

TrackLib And Legal Sampling

SPEAKER_01 2:14

Okay. So for anybody that's not familiar with them, there is a company called TrackLib, right? Now, TrackLib are a platform that enables producers and artists to sample music legally. So you find a song to sample, kind of like a digital crate digging exercise, right? But instead of having to then, you know, uh sample something and then figure out if you're going to clear it, they flip it the other way around. So everything that's in that part is already cleared. Now, you'd find a song to sample, you create music using those original recordings, you clear the samples before the release, and then the track ownership and provenance, i.e., where it came from to begin with, the intangible on the song and the recording is recognized, and then it works to compensate rights holders. Now, why can't that work in AI? Well, there's probably a few reasons, and I can think of a few and I'll tell you what they are. But before then, let me present this to

Turning That Model Into AI Licensing

SPEAKER_01 3:06

you. If we applied this to AI generative systems, this is what it could look like. Licensed training models, licensed training catalogue, so AI companies train only on music where labels, publishers, artists, and right holders actually opt in. Training provenance database, so creators can see whether their music was actually included in the process of training, revenue participation, right? So artists whose catalogues contributed to training material receive royalties or usage-based payouts, or if it's like track lib, then actually both. And then there is influence reporting. So AI systems could disclose the training domains or stylistic influences on some level. There is also then the clear commercial rights, so users generating music know exactly what they can legally release. So if we were to put that into a little chain, this is what the chain could actually look like. Artists upload their catalogue, chooses training permissions, AI company licenses access, the model trains, then generated outputs create revenue. Rights holders receive compensation according to agreed terms. Now, this is a little bit like what tracker do. And again, I'm using them as a kind of example here because they've kind of made this work in sampling and it works really, really well. There's been hit records with it, and it works fairly and it works transparently. Now I know it's a different beast when it comes to AI because there's so many other levels to it, but I think we can learn something from it. Now let's just simplify it into this. If we look at two different pie charts, one for the song and one for the recording, then we can see that actually a little piece of both of those two things will be eaten into through training data. So, in other words, if we've trained on this song over here, then we've used kind of like some influence from its recording, then that might be right. In order for us to use this in this song, we have to give away, let's say, 12% of our song or our music publishing side of things, and then we have to pay this fee over here for the a micro fee, right, for the use of the recording. Now, if that were to happen, that happens in the sampling world in tracklib. If that were to happen in the AI world, I know it's more complex, but it's still possible. However, there are problems that

Scale Attribution And Other Hard Bits

SPEAKER_01 5:26

come with it. First of all, there's a major scale difference, right? Track lib licenses humans, right, to deliberately sample something and AI models ingest millions of words statistically rather than just reproducing single pieces. That's one problem, but I'm sure somebody could build something to overcome that. Value allocation. So if an AI trained on 10 million songs to generate one output, i.e. one song, how do you decide who gets paid? Now that is difficult, but at the same time, it could be that actually that could be filtered through by looking at its provenance, by looking at its characteristics, which is not too dissimilar than the technologies used by, let's say, Shazam when they scan a waveform or by Spotify when they do raw file analysis to understand the characteristics and elements of the songs recording. So if that similar kind of thing was applied to this, it might be possible to track that and then remunerate them further down the line. Now, influence versus copying. AI models usually learn patterns rather than storing reusable samples. So making compensation accounting a lot harder. Competitive pressure also, so licensing everything up front is expensive and slower than scraping public internet data. Of course, it's gonna be, it's gonna be harder. I'm not saying this is an easy transition to do, I'm just saying I think it might be possible. One thing's for certain, it would create a it would reduce a lot of industry tension. Now there is proof of concept on this already. So

Proof Of Concept From Real Companies

SPEAKER_01 6:58

we're gonna look at two different examples where we're kind of starting to see this thing happening already. So Clay Vision, who are a large music model, are particularly interesting here because it publicly claims that large music model is trained exclusively on licensed content and it has secured agreements with Universal, Warner, and Sony to be able to do that. Also, Eleven Labs Music are a very interesting company, they license training material from Merlin and Cobalt for its music system instead of relying on unlicensed major label data. So we've already got a couple of examples there where it's already started to happen, but a lot of people don't necessarily know about this yet.

From Scrape First To Licence First

SPEAKER_01 7:40

Now, traditionally speaking, the music industry has often been slow to react to problems. We've seen that in the past, and we don't really want to see it again with this, although we might already be there a little bit, but it's certainly not too late. There is often a culture of scrape first, litigate later, but I think that the music industry is now starting to turn into licensed catalogues and partnership models, and that again would fall in line with what I'm projecting here today. So there you have it, a little summary. Now it might be that some people who are watching this or listening to me right now, uh, people that I know of in the music industry going, well, Johnny, you know what? That's lovely, but it's not that simple. I'm sure it's not. I'm not projecting for one minute that this is a simple transition. But the problem is it might be too late if we leave it too long. And this is one thing that could shift things around a little bit. Now, I know that there's still very much a moral issue for a lot of music creators, but I also know that there are many, many, many people now using AI to create a part of something that they're working on, and that part then might get then re-recorded later down the line. Now, AI usage is not going to be able to track that, is it? That's something that we at the moment we're in a declaration model. We have to say, oh, we've used this, and therefore we we we say that because we've used this, then uh you have to print it and say that we've declared usage on this. That's lovely and it's a step in the right direction, but it's also very flawed because it relies totally upon honesty. And I'm not saying people aren't honest, but you know, sometimes people some people aren't, right? And that's difficult as well. So actually, if the menu was license something for training and then partner it correctly, then we're moving away from that let's sample this, let's litigate later, which is kind of one of the problems of the past in the recorded music industry, especially in the 20th century. Sampling in itself is still a very difficult thing for some people to understand because it's like history building on history. And if you don't sample music or you are not a you're not open to being sampled, that's okay. You don't have to be because everything that I've projected here, um, you can opt out of it, anyone can. But if you opt into it, it just creates an entirely new economic landscape for rights holders, for independent music creators, for artists, for producers, for songwriters, for performers, all of the people that we know in the music industry that are at the heartbeat, the lifeblood of music, those that are making music, right? If those people could be remunerated if their if their work is used as training data, then it can only be a good thing for the economics going forward. And that is the idea of today's podcast.

Your Take And The Future Of AI

SPEAKER_01 10:35

Now, let me know what you think, guys. Am I off on something here that you think, crikey, me, he's lost it? What's he talking about here? Or are you thinking, Crikey, he might be onto something here? There's already proof that this is moving in that direction. Could it change completely? Could a big tech company all of a sudden go bang? Actually, we're going to be completely transparent about our training model and go license model first and then make it accessible to people, or is there going to be a new company that comes around that changes everything forever? That does this first? Only time will tell. Anyway, that's enough from me today. Have a great day, and may the force be with you.