75,000 AI songs a day on Deezer. The listener, somehow, is fine.

Last week, Deezer’s CEO Alexis Lanternier published a number that should — by every reasonable instinct — be terrifying for working songwriters.

75,000 AI-generated tracks are now uploaded to Deezer every single day. That’s 44% of all new music being added to the platform. Up from 10,000 tracks a day in January 2025. A 7.5× growth in twelve months. Two million AI tracks a month. A hundred thousand a week. Three a second, while you’re reading this.

If you write songs for a living — or even for the hope of one — your first reaction is probably the obvious one. That’s the end of us. The flood is here. We’ve been replaced. And on the platform-economics side, there’s a real argument for that anxiety, which I’ll get to.

But there’s a second number buried in the same report that almost no one is repeating, and I think it’s the more important one.

Those 75,000 AI tracks per day, all 44% of new uploads, account for only 1 to 3% of actual streams.

Sit with that for a second. Almost half of everything new being uploaded to one of the world’s biggest streaming platforms is AI-generated. And almost nobody is choosing to listen to any of it.

This post is about why that gap exists, what it tells us about listening and creativity and what songwriting actually is, and what working songwriters should — and shouldn’t — do about any of it. It’s a long one. The question deserves a long one.

The asymmetry nobody is talking about

This is the most interesting fact in the whole story, and it’s getting buried because it doesn’t fit either of the two narratives the music press wants to tell — AI is taking over, panic now, or AI is no big deal, calm down. The truth is stranger than both. AI music is winning the upload war and losing the consumption war by an enormous margin.

If 44% of supply produces 2% of demand, the average AI track on Deezer is being streamed something in the order of 30 to 50 times less than the average human track, per upload. The platform is full of AI music. The listeners are uninterested.

Then the third number lands. Deezer says 85% of AI music streams are flagged as fraudulent and demonetised — meaning they’re not real listening behaviour at all, they’re bot farms hammering AI tracks to siphon money out of the streaming royalty pool. So even that 1-to-3% of plays is mostly not human ears. The actual proportion of AI music being chosen and listened to by real people is somewhere between very small and vanishingly small.

To put it in perspective: in 2025, Deezer detected and tagged 13.4 million AI tracks. The platform’s total catalogue contains roughly 110 million songs. AI music is now more than 10% of everything Deezer hosts, and growing by 75,000 tracks a day. And listeners, by their actual behaviour, are choosing it less than two percent of the time.

The flood is here. The listener, somehow, is fine.

How we got here

Two timelines collided to produce this moment, and it helps to see them both before we talk about what to do.

The first timeline is streaming economics. Spotify, Apple, Deezer, Tidal, Amazon — all of them pay royalties out of a pool that’s divided pro-rata across total streams. Spotify pays out of total subscription and ad revenue; the more streams happen on the platform, the smaller each stream’s share of the pool becomes. This was already a problem before AI showed up. A song that earned a writer $5 in 2014 might earn $1.50 today, not because the song is worth less, but because the supply side of the streaming equation has expanded faster than the demand side. More songs, flatter pool.

The second timeline is generative audio models. Suno, Udio, Stable Audio, the open-source equivalents — by 2024 the technology had matured to the point where a competent-sounding three-minute song could be generated in under a minute, for a fraction of a cent of compute. By 2025, the tools were available in browsers, on phones, free or near-free. Anyone with a laptop and a credit card could push hundreds of “songs” a day to a distributor like DistroKid or CD Baby and have them end up on every major streaming platform.

The streaming model rewards volume. The generative model produces volume. The bot-farm fraud economy connects them: pay a few hundred dollars for a few hundred AI tracks plus a few hundred fake listeners, and the streaming royalty math turns it into a profit, however small, multiplied across thousands of accounts.

That’s how a platform ends up with 75,000 AI uploads a day. None of that is mysterious. It’s the predictable outcome of two systems — pro-rata streaming royalties and free generative audio — pointed at each other without a filter in between. The platforms are now building the filter in real time. That’s most of what the Deezer announcement actually was.

What AI music structurally is, and isn’t

It helps to be precise about what these tools actually do, because the public conversation tends to slide between AI generates music and AI writes songs, and those are very different claims.

A generative audio model — Suno, say — is trained on millions of existing recordings. Internally, it’s learning a statistical map of what “music” sounds like — the distribution of frequencies, the typical relationships between chords, the way a vocal phrase usually ends, the way a snare lands relative to a kick. When you give it a prompt (“country song about a bad breakup, female vocal”), it draws from that distribution and produces an output that fits the statistical shape of the prompt. The output is, by construction, an average. It’s plausible because it’s drawn from the centre of the data. It’s bland because that’s where averages live.

What AI generation can do, very well, is produce music-shaped artifacts. A track that has a verse, a chorus, a bridge in the right place, a vocal that sits in tune, a mix that sits in the right loudness. From three feet away, indistinguishable from a song. Do not underestimate this. The bar for “passable music” has been crossed. Any argument that depends on AI music sounding bad is going to age badly.

What AI generation cannot do — and this is structural, not a temporary limitation — is have an interior life that the song is metabolising. A real song is, almost always, a working-out. The writer started with something they couldn’t say plainly, and the song is the form that thing finally took. The bridge changed because the writer’s understanding of their own grief changed in the third week of writing. The line about the hotel room is in there because the writer was actually in the hotel room and the lamp had a green shade. The chorus rhymes the way it does because the writer rejected six other rhymes that didn’t carry the weight of what they meant.

A generative model has none of that. It can produce a song about a hotel room with a green lamp if you prompt it to, but the green lamp is ornament, not source. It’s reverse-engineered from the language of songs that already exist about hotel rooms. There is no person whose specific hotel room it was. The song is a shape, not a working-out.

This isn’t a moral claim. It’s a structural one. And it’s why the Deezer numbers look the way they do.

Why the listener is fine, and will stay fine

Listening is not what the AI doom-loop assumes it is. The doom loop assumes a song is a thing — a file, a product, an artefact — and that if you can generate a thing that’s technically indistinguishable from a human song, you’ve replaced the human. But that’s not what listening actually is.

Listening is a relationship. When a human listens to a human song, they’re not just processing audio. They’re connecting to another person’s specific moment — the grief, the bridge written in a hotel room, the I-meant-this of the whole thing. That intentionality is half of what music is. The other half is the sound. AI gets the sound. It cannot fake the intentionality, because there’s no one home behind the file.

You can argue whether listeners consciously know this. I’d say they don’t have to. The way a song lands in a body is partly a transmission of meaning from one human to another, and when the sender is empty, the transmission is weaker. Listeners feel it before they can name it. That’s the difference between a song that makes you cry on the motorway and a song that’s perfectly arranged but somehow does nothing to you.

There’s an experiment people do online where they play a clip of an AI-generated song and a human song side by side and ask listeners to guess which is which. Listeners get it right far more often than chance, but not always. The “always” part isn’t the point. The point is that even when they can’t articulate what’s wrong, they consistently say things like the AI one feels weirdly empty, or I can’t tell why but the human one stays with me longer. They’re not detecting the production. They’re detecting the absence of a sender.

The Deezer data is exactly what I’d expect that hypothesis to look like in numbers. Not zero AI plays — some tracks are good enough to function as background, and most listening at most times is half-attention background anyway. But not parity, not anywhere close. The instinct to seek out a person you can hear meaning something is still wired into the species, and it’s still doing its job.

The thing that makes me confident this won’t change with better models is that the gap isn’t a quality gap — it’s a category gap. Better AI won’t make it close to parity any more than better photo-realism made the public stop preferring real photographs of people they love over AI-generated ones of strangers. The relationship is the point. The artefact is downstream of the relationship.

The two real problems for songwriters

So if listeners are mostly fine, what’s the actual grievance? Two things, and they’re both worth being angry about.

One: the platform economics. Streaming royalties on every major service are paid out of a pool, and the pool is divided pro-rata across all streams. Even when 85% of AI streams get demonetised, the remaining 15% still take a cut from the same pool that would otherwise be paying real songwriters. And every track uploaded — AI or human — adds to the supply side of the equation, which dilutes the per-stream payout for everyone, even when nobody is actually listening to most of those new uploads. The flood doesn’t take over the listener’s attention. It does take some of the money that would have gone to real writers, even when listeners aren’t choosing it.

That’s a real grievance. It’s also worth knowing that the platforms are now incentivised to fight it — not because they love songwriters but because the fraud bleeds their margins. The Deezer announcement is what that looks like: detection systems, demonetisation, AI tagging, licensing the detection tech to PROs and collection societies. Spotify recently started removing tracks from artist profiles that fail their authenticity checks. Apple Music is rumoured to be doing the same. Strange bedfellows, but on this particular fight, the platforms are on the same side as the songwriter. Use that.

Two: the discovery problem. Even if AI tracks aren’t getting listens, they’re occupying shelf space. The algorithms have to wade through them. Editorial curators have to filter them. Every search result, every “related artist” suggestion, every algorithmic playlist slot is now competing with infinite supply that didn’t exist five years ago. For an unsigned songwriter trying to break in, the noise floor just got louder. You don’t lose listeners to AI; you lose visibility through it.

The reasonably good news on this front: listeners are evolving filters in real time. The same way every person under thirty instinctively knows whether a YouTube channel is human or AI within about five seconds — there are tells, you can feel them — listeners will develop the same instinct for music. The platforms are already helping. Deezer is labelling AI tracks. Spotify will follow. The labels will get more aggressive about authentication because consumer trust depends on it. The discovery problem is real, but it’s a 2026-2027 problem with a relatively short half-life. By 2028, AI tracks will mostly be quarantined into their own taxonomy and the human catalogue will have its own clean shelf again. The job between now and then is to be audibly human enough that the filtering process works in your favour.

What is creativity actually for?

Here’s the philosophical question that hides under all of this, and the one I think songwriters should think about most carefully. What is the act of writing a song actually doing? What’s the function of it, in a human life and in a culture?

If a song is a product — a unit of entertainment that can be substituted for any other unit of entertainment of equivalent quality — then yes, AI is going to compete with it on those terms and probably win, because the marginal cost of an AI song is approaching zero. If this is the frame you’re using, the panic is rational.

But songs have very rarely been products in any deep sense. Songs are the way one person tells another person what it is to be alive. That’s not me being romantic about it. That’s literally how the form works. “Hey Jude” is a song Paul McCartney wrote to comfort John Lennon’s son Julian during his parents’ divorce. “Cover Me Up” is a song Jason Isbell wrote to his then-girlfriend, now wife, the night he got out of rehab. “Kyoto” is a song Phoebe Bridgers wrote about feeling dissociated in the middle of a tour she’d dreamed about her whole life. “The River” is a song Bruce Springsteen wrote about his sister’s marriage falling apart in industrial New Jersey. “Hurt” is a Trent Reznor song about heroin addiction; Johnny Cash sang it three years before he died and it became about something different, about a man at the end of his life looking back. The songs people actually love are nearly always the ones you can trace back to a particular person, in a particular moment, trying to work out something that mattered to them.

That’s what creativity is for. It’s not decoration. It’s not content. It’s the way humans process experience that’s too complicated to process any other way. We sing because there are things we can’t say plainly — love, grief, hope, regret, the specific terror of being twenty-three — and the song is the technology we built to handle those things together. The listener participates because they’ve also had things they couldn’t say plainly, and they recognise their own experience in your working-out of yours.

You cannot generate that. Not because the technology isn’t good enough, but because the function isn’t what the technology is doing. A generative model is producing music-shaped artefacts that the public mostly experiences as content. Songwriting at its best produces shared metabolisations of experience. These are different categories. Even when the artefacts look similar, they’re playing different games.

The Deezer numbers are the listener telling us, in aggregate, that they already know which game they’re in.

Specificity as the only moat that matters

This is the practical takeaway, the one I’d want every songwriter reading this to leave with.

The songs AI struggles to compete with — and the songs the next decade is going to reward most — are the maximally specific ones. The ones that could only have come from one person, in one place, at one moment.

Consider:

Now Mary’s dress waves / Mary’s dress sways / Like a wave on a sea that the wind sends and sweeps away

That’s Springsteen, Thunder Road. Mary’s dress. Not “her dress.” Mary’s. A specific name attached to a specific dress moving in a specific way. AI can produce a song about a girl in a dress. It cannot produce Mary. The specificity is the song.

I’d say I’m doing well, I’m a kid in a candy store / I picture the show in my head and there I am, hoarse from screaming

That’s Phoebe Bridgers, Kyoto. The narrator is in a candy store metaphor and on the other side of the world from where she imagined herself, and she’s not winning. AI can produce a song about being homesick on tour. It cannot produce Kyoto, because Kyoto is the specific shape of one person’s specific dissociation in one specific city. Take the proper noun out and the song is gone.

I sobered up and I swore off that stuff / Forever this time

Jason Isbell, Cover Me Up. The line works because of the forever this time — the qualifier that admits all the other times that weren’t forever. That’s what real recovery sounds like. A model trained on recovery songs cannot produce forever this time unless you prompt it to, and even then it will be ornament, not the thing itself.

The pattern is the same in country, indie, hip-hop, R&B, folk, pop. The songs that age, the songs that get covered, the songs that survive fashion are the songs where the writer named something specific enough that the listener could see through it to their own specific thing. The Mary. The Kyoto. The forever this time.

This is what every songwriter teacher has been saying for decades, and it was always true, but it was also implicit. Before the AI flood, the human-ness of a song was a free baseline. Specificity was a craft preference, but you could still get away with vague — baby, love, tonight, forever — because at least there was a person behind the vagueness. Now the vagueness has competition that costs nothing to produce. The Mary doesn’t.

That’s the entire shift. Specificity used to be a stylistic choice. It’s now the only competitive moat that matters.

The practice that follows

If the value of human songwriting has gone up, not down, then the practice that follows is the opposite of the panicky one. Don’t write more, faster, cheaper to compete with the volume. You can’t, and you shouldn’t try. Don’t try to be undetectable from AI. The whole point is that you shouldn’t be.

Write fewer songs, more carefully.

Put the specific detail in. The wrong street name. The actual time of day. The real person, even if you change their name in the second draft. The thing your mother said to you that you’ve been carrying around for fifteen years. The way the light came through the window of that one flat in Galway in March 2018. Not the universal version — the that version. The one only you can write because it only happened to you.

Sing it like it’s the only chance you’ll get. Be a person, audibly, in the room — because that’s the one thing the flood can’t fake, and the listener can hear it from the first bar even when they can’t articulate why.

Forget about the algorithm. Spotify’s algorithm and TikTok’s algorithm are going to collapse under the weight of synthetic content within the next two years. The writers who survive that collapse will be the ones whose audience is with them, not just exposed to them via algorithmic chance. Build that audience by being unfakeably yourself, song by song, and by writing things only you could have written.

Keep your community. The most resilient songwriters in the next decade will be the ones with email lists, supporters, regulars at venues, people who know their actual name. AI is going to commoditise the file; it cannot commoditise the relationship. The relationship is what streaming has been quietly under-valuing for fifteen years, and it’s about to be the most valuable thing in music again.

Write better split sheets. (Yes, really.) When AI fraud and pool dilution are eating into per-stream payouts, the writers who lose the least are the ones whose splits, registrations, and admin are airtight on the upside. If you write a song that does get heard, you want every penny of it ending up in the right pocket. That’s been true forever; it just got more true.

The whole story

The flood is real. 75,000 AI songs a day is not a typo. By next year it will probably be 150,000.

The listener is also real, and the listener is fine. They are, in aggregate, choosing to listen to AI-generated music something between 1 and 2% of the time, and most of that 1-to-2% is actually fraud bots, not human ears. The remaining real-human AI listening is probably under half a percent of total streams. That figure may inch upward; it is not going to become the majority. The relationship is the point. The artefact is downstream.

The grievances are real. Pool dilution is real. The discovery noise floor got louder. Both are temporary problems with active solutions, and the platforms are mostly on your side, however reluctantly.

The opportunity is also real, and it’s bigger than the threat. Human songwriting just had its scarcity declared, in public, by the numbers. Specificity — your particular life, your particular voice, your particular Mary — is now the moat. The writers who lean into that are going to have a better decade than the ones who try to compete with machines on machine terms.

Write the song you’d be embarrassed to send to anyone. Sing it like it’s the only chance you’ll get. Be a person, audibly, in the room.

75,000 AI songs a day. None of them yours.

That’s the whole story.

Written by Daniel Joseph Healy. If this was useful, the tools page has more like it, and the homepage has a newsletter signup.