I’m Optimistic the New AI Deals Will Be Good for Artists. But What If I’m Wrong?

There’s a dystopian future ahead for artists & the industry if things break a certain way.

As I’m sure you read in my last post, labels and Gen AI companies are now playing nice in the sandbox after a year and a half of lawsuits and the predictable media posturing. Suno, Udio, ElevenLabs, and other emerging players have entered into agreements with labels and publishers that - on the surface - seem to establish the contours of a responsible future for music and AI.

We got the settlements. We got the licensing. We got the commitments to deprecate unlicensed models and rebuild them on sanctioned data.

It finally feels as though the industry has turned a corner. But have we?

My optimism deserves some scrutiny. Yes, the music industry’s relationship to technology is light years better than it was when streaming started (don’t even mention file sharing). Yes, a lot of the issues we have in music (too much of it, streaming fraud) aren’t because of AI - yet get conflated with it.

Because if we zoom out - and examine not just the immediate legal victories but the structural forces now in motion - we can game theory a very different outcome.

One in which these deals, despite their good intentions and short-term gains, end up accelerating changes that are deeply unfavorable to artists and the music ecosystem.

Why Independent Artists Could Lose Most (The Inverted Power of Training Data)

One of the most overlooked dynamics in Gen AI is what I call the “inversion of market share.” In the traditional music business, major labels dominate revenue. They have the catalogs, the marketing muscle, the global distribution, and therefore the leverage. But when it comes to training data, the majors represent only a sliver of the total universe of recorded music. The vast majority comes from the independent sector - millions of artists distributing through TuneCore, DistroKid, CD Baby, and others.

Thus, independent artists have contributed most of the music that trained early AI models, yet they are not the ones at the negotiating table. Labels are securing licensing deals, collecting settlement payments, and dictating the next wave of AI governance. And yes, although Merlin is part of this - what about the millions of artists in the long tail who supplied the bulk of the training data? These are the artists that are not earning streaming revenue, but their music can form the secret sauce in Suno’s $2.45B valuation? What protections or compensation will they receive?

The Problem of “Too Much Music” (Accelerated by AI)

Before Gen AI arrived, the industry was already buckling under oversupply. When I was Director of Artist Partnerships and Industry Relations at Pandora in 2015, we had about 18 million songs in our library. A decade later, digital music recognition firm Audible Magic estimated that number at more than 220 million - a ten-fold jump. Streaming platforms have been straining under this load, with algorithms in constant triage mode and discovery collapsing for all but the most established acts.

Then came AI, adding millions of tracks in just the last two years - and with the potential to add infinite ones. Human artists have always competed with each other for attention; now they’re competing with machines that can generate 10,000 tracks in the time it takes a bedroom producer to open Ableton.

If today we have too much music, AI pushes us into way too much music. And because recommendation systems are already optimized for volume and efficiency, they’ll naturally gravitate toward whatever is cheapest, most abundant, and most predictable. In an AI-saturated world, that could quietly move the system toward the safest, blandest kinds of creativity.

I hope better curation, richer metadata, or artist-centered platforms can counterbalance this. But what if they don’t?

What if oversupply becomes so overwhelming that emerging human-made music simply can’t break through at all?

Frictionless Creation (The Loss of Cultural Scarcity)

One of my favorite quotes comes from the late producer Steve Albini. In response to a studio engineer complaining about a finicky reel-to-reel tape machine, Albini reportedly commented that recording equipment being hard to use was a benefit, because it “keeps the riff-raff out.”

For all of technological musical history, creativity contained *some* friction. Writing a song required time, skill, emotional energy, and craft. Recording required resources. Releasing music required some idea of how it all worked.

Gen AI vaporizes those constraints. The distance between ideation and creation has collapsed. If a thought can be turned into a finished track instantly, creation becomes frictionless:

• Scarcity disappears

• The symbolic value of craft diminishes

• The distinction between inspiration and production gets muddied

• The personal identity of “an artist” becomes harder to sustain economically - and psychologically

This is not to say frictionless creativity is bad for everyone. It undoubtedly empowers hobbyists, educators, marketers, and many working musicians. But the scale at which AI can produce content threatens to swallow us all.

Then what if copyright offices around the world decide to give Gen AI publishing rights? This machine music (plug for my book!) - essentially - would have the same commercial benefits as human music. Would a child who loves listening to music be incentivized to try and create it?

The Rise of Synthetic Data (and the Collapse of Attribution)

Perhaps the most under-discussed piece of this whole debate is synthetic data: The AI-generated music that’s then fed back into train the next generation of models.

When labels and artists ask me whether they should license their music to AI companies, my first answer is always: that’s a personal decision. My second: if you’re inclined to do it, do it now.

Look at what’s already happening in text, images, and video: synthetic data has become a powerful accelerant for training. We should not kid ourselves about how quickly this technology is improving. Inference scale (the amount of data and the number of iterations a model needs to hit the “right” output) is dropping fast. Back in the dark ages of Gen AI (late 2024), you might have needed a million songs for a certain level of quality; now you might get there with 100,000. Soon, it’ll be less.

And if a model needs variation in tempo, key, or genre? You don’t need a diverse human catalog when you can speed up, transpose, style-transfer, and timbre-transfer existing recordings. One song can be spun out into 10,000 variants across tempos, keys, and styles.

For a while, AI researchers worried about “model collapse” - the idea that if models trained too heavily on their own outputs, everything would converge into the same bland mush. Increasingly, though, that risk looks manageable or at least solvable.

Which could lead to an anti-artist conclusion: if an AI company can build a high-quality model off a stock pile of library music and its synthetic derivatives, the question of human attribution starts to look irrelevant. These models simply won’t need a label - or artist’s - IP.

My hope is that synthetic data remains supplemental rather than foundational. But if I’m wrong, synthetic data could become the mechanism by which human artists permanently lose leverage in negotiations with AI companies.

Deprecating Models Doesn’t Make Sense (So Why Is It Happening?)

A curious feature of recent AI-label settlements is the promised “deprecation” of unlicensed models. On paper, this feels like accountability: labs acknowledging they used unlicensed data and committing to rebuild their systems from scratch.

But from a technical perspective, this is implausible. Deprecating a model trained on a rich, heterogeneous dataset - then retraining it on a tiny subset of licensed major-label recordings - would produce an inferior system. AI models rely on massive diversity to learn generalizable musical structure. A narrower dataset is a weaker dataset.

So why deprecate? One possibility is technology transfer. Maybe Udio does not want to give labels insight into architectures, generation pipelines, or synthetic data workflows that allow them to build their own internal AI engines - effectively giving UMG a way to make a top-flight AI model. Another possibility is that both sides simply need a cooling off period while new norms form.

But if I’m wrong in my optimism and deprecation is merely a PR exercise, then we may be entering a phase where compliance is purely symbolic rather than substantive - and where the true long-term trajectory is toward models increasingly insulated from human oversight.

It wouldn’t be the first time a tech company ignored the rules.

Major Labels as AI Models (Where’s Your Future?)

Speaking of UMG building an AI model: The economic incentives for labels in an AI-driven environment are straightforward. Streaming rewards volume, not quality. Major labels are already heavily invested in production music, mood playlists, and catalog strategies designed to maximize passive listening revenue.

If Gen AI gives labels the ability to produce unlimited quantities of high-margin functional music, then becoming an AI-powered content factory is a rational business decision.

Labels could theoretically maintain two parallel lines of business:

1. Human artistry, focused on fandom, touring, and brand partnerships.

2. AI-generated catalogs, optimized for sleep playlists, study music, workout soundtracks, and mood-based listening - on DSPs and elsewhere.

And given current A&R costs and royalty structures, the latter could become more profitable than the former.

If I’m wrong, and labels use AI only to enhance human creativity, then these concerns dissipate.

But if I’m right - or even half-right - we may see an industry where human artists are no longer the primary revenue drivers, but one category among many in a vast, synthetic content marketplace.

Back to the Future of Music Consumption: Fans vs. Listeners

Before streaming, the music economy bifurcated cleanly:

• Fans bought CDs and downloads

• Listeners consumed radio

Streaming collapsed those two groups onto one platform. But AI threatens to split them again. Gen AI music is extraordinarily good at functional listening: background audio for work, exercise, sleep, or study.

If consumers gravitate toward AI-native environments for passive listening, human artists will be pushed back into the sharply limited “fandom economy,” where only a small percentage can earn sustainable income.

I hope that AI strengthens fandom and creates new forms of engagement (which I have written about before). But what if AI simply becomes the new radio - not a driver of discovery, but a replacement for it?

Optimism Requires Preparing (For the Hard Questions)

I’m still optimistic that these new licensing deals between labels and AI companies can end up benefiting artists. There are some very real forces working in our favor: fans, almost universally, do not want to listen to fully Gen AI music; we still place a premium on human expression. Culturally, we want to nurture the creation of art and reward the businesses that help it flourish. And as overwhelming as these tools can feel in their speed and scale, they’re not entirely alien - they’re part of a long line of technologies, from polyphonic synths to DAWs, that were once feared and are now just part of how we make music.

But if the dynamics outlined above unfold unchecked - oversupply, synthetic data, loss of attribution, independent-artist marginalization, frictionless creation, and labels prioritizing AI models - then these recent deals may not be the victory they appear to be. They may instead be the opening act of a profound restructuring of the music ecosystem, one in which human artists must work harder than ever to be heard above the noise: not of their peers, but of the machines trained on their work.

I hope I’m wrong. But we owe it to artists to take seriously the possibility that I’m not.

---

I have a book coming out in February on Routledge. It’s called “Machine Music: How AI is Transforming Music’s Next Act.” You can learn more and sign up to preorder here:

https://machine-music.ai/

---

If you find value in this Substack, please consider a paid subscription. Proceeds will go to:

• Keychange U.S., an organization that seeks to support and advocate for underrepresented genres in the music industry.

• Why Hunger, which connects those who are food insecure to nutritional resources.