How much of Spotify is AI music?

First published 11 May 2026

And where is it heading?

Every platform shift in music has changed how we consume it. This one is different because the platform has picked the flood's side.

Nobody knows, because Spotify won't say. Deezer will: by June 2026 more than half the tracks uploaded to it each day were fully AI-generated, around 90,000 a day, up from 28% the previous September. Those tracks earned 1 to 3% of streams, and Deezer says up to 85% of those streams were fraudulent. The flood is real, the audience isn't, and the platform has picked the flood's side.

This disparity between supply and demand is where it gets interesting.

Napster changed how we owned music. iTunes changed how we paid for it. Spotify changed how we found it. This one changes what counts as music in the first place.

Spotify recently came out and said they removed 75 million spammy AI tracks in the last twelve months, which sounds like enforcement until you notice what's being enforced.

In essence, the embarrassing AI music gets removed. The competent AI music (the kind that racks up a million monthly listeners under a synthetic band name without anyone complaining) is left exactly where it is. That's like a pub barring its worst-behaved drinkers every night and calling it a drinking policy. It's not tackling the problem. It's protecting the licence.

The conventional take is that AI music is bad and listeners can tell. However, according to research, 97% of them actually can't. The consequences of this fact are worth tracing downstream. That's what I've done here.

Below: here and now, imminent, on the horizon, the long view, and five behavioural shifts running underneath.

01. Here and Now

Already landed, observable today

Most coverage talks about AI music as a future problem. The royalty pool is already diluted. Synthetic acts are already on playlists. Spotify is already farming spam at industrial scale. The damage has started.

• The royalty pool is already being diluted — Spotify pays out a fixed share of revenue divided by streamshare. Every AI stream takes from the same pool that pays working musicians. 0.5% sounds small until you remember it's 0.5% of every artist's income.
  • The 1,000-stream threshold raised the bar — Spotify introduced a minimum of 1,000 streams in 12 months to earn any royalties at all. AI uploads make that threshold harder to clear, because the algorithmic oxygen now gets eaten by 50,000 daily AI tracks competing for the same playlist slots.
  • It mostly hurts the marginal earner — Taylor Swift is fine. The artist earning £200 a month who now earns £0 because the cutoff got harder is the actual loser. The harm is concentrated at the bottom of the distribution, which makes it easy to ignore.
• Synthetic acts are already on the playlists — The Velvet Sundown amassed over a million monthly listeners as a fully AI-generated band before admitting what it was. People liked it. The harm wasn't in the listening. It was in the disclosure.
  • Listeners felt cheated retrospectively — Spotify's own exec said the news cycle would have been very different if the band had labelled itself from the start. Which is another way of saying: the music was fine until the label changed.
• Spotify is already running a fraud farm it pretends to police — 75 million spam tracks removed in twelve months. That is not a moderation problem. That is a sign the platform has become a vehicle for royalty fraud, where the music is incidental to the streaming maths.
  • The fraud loop is simple and the platform knows it — Generate a track in Suno. Upload it. Get bots to stream it past the 1,000-stream line. Collect royalties. Repeat at scale. The 75 million number is what got caught.

02. Imminent

Visible to those watching, not yet at scale

Visible to those watching, not yet at scale. Functional genres are being eaten first. Discovery surfaces are tilting. The middle of the industry is being hollowed in slow motion. The next twelve months make these obvious to anyone paying attention.

• Functional genres are being eaten first — Lo-fi study beats. Sleep music. Yoga playlists. Ambient. Library music for podcasts. AI is eating the genres where the listener brief was already 'audio wallpaper, ideally cheap'. The genres being colonised first are the ones nobody was emotionally invested in.
  • Most of this was already content, not music — The sleep playlist made by an anonymous producer for a content mill in 2019 is not categorically different from the one Suno makes in 2026. AI is doing more efficiently what humans were already doing on a piecework basis.
  • Indie filmmakers and podcasters got cheap soundtrack access — Library music used to cost real money. A YouTuber, a small podcaster, an indie filmmaker can now licence usable backing tracks for almost nothing. The same flood that's a red node for library composers is a green node for the people who used to be priced out.
• Discovery surfaces are tilting toward retention metrics — Discover Weekly used to feel like a curated act of taste. AI tracks are engineered for the first 30 seconds because that's where the skip threshold sits. The recommendation surface increasingly rewards what doesn't get skipped, not what gets loved.
  • The algorithm was always doing this — AI music isn't introducing a new problem. It's exposing that Spotify's recommendation system was optimised for retention all along. AI just makes it cheap to game the metric the platform was secretly using.
• The middle of the industry is being hollowed — Session musicians. Library composers. Jingle writers. Sync licensors. The unglamorous middle that made a living from functional commissions. AI doesn't eat the music industry uniformly. It eats the middle first.
  • These people aren't on the cover of Rolling Stone — The £40k-a-year session player who pays a mortgage by playing on adverts and sync placements doesn't make headlines when their work dries up. Which is part of why the hollowing is going unremarked while everyone watches the Drake of it all.

03. On the Horizon

One to three years out; the forks are real and being reached

One to three years out, the shape of things gets set. Lawsuits get decided. Disclosure rules get drafted. Labels pick sides. Spotify positions itself for a market that doesn't quite exist yet. The forks are real and they're being reached.

• Mid-tier indie pop is next on the menu — The production template for mid-tier streaming pop is already so standardised it's nearly procedural. The genre that fills 'chill vibes' and 'sad girl autumn' playlists is the next functional category, even though it pretends to be expressive.
• Spotify restructures its label relationships — Every AI stream is one Spotify doesn't pay Universal, Sony or Warner for. AI is the first content category in Spotify's history where the platform could plausibly own the supply chain. They will not waste that.
  • Söderström's 'not our decision' line is positioning, not neutrality — When Spotify's product chief says it shouldn't be deciding what tools artists use, what he means is: please don't make us choose, because the maths is already choosing for us.
• Disclosure regulation forces a label — The EU and the UK are both drafting AI-music disclosure rules. If listeners have to be told before pressing play, the Velvet Sundown effect becomes the norm: people enjoy until they're told, then they're annoyed. Whether that changes consumption is unclear.
• A major label pulls catalogue and triggers a renegotiation — Universal has form for this with TikTok. If a major decides Spotify's AI tolerance is a contract breach, the platform faces a choice between its biggest suppliers and its highest-margin content. They'll pick the suppliers. Until they don't.
• Suno or Udio loses a defining lawsuit — Both are being sued by the major labels over training data. A loss doesn't kill AI music, but it changes who can legally generate it and at what cost. The flood would slow. The economics would shift.

04. The Long View

Three to ten years out; the dust settles into a pattern

Three to ten years out, the dust settles into a pattern. Listening becomes background. The platform becomes the producer. The shape of what gets called 'music' on streaming services in 2035 is being decided now, mostly by people who aren't musicians.

• Listener taste atrophies as background consumption dominates — The 97% who can't tell AI from human aren't being fooled. They're not paying attention. Music as foreground experience is losing ground to music as ambient mood, and AI is uniquely suited to making ambient mood at industrial scale.
• Spotify generates the music it streams — Söderström has already hinted at 'AI derivatives' as a revenue stream. The end state where Spotify generates the music it streams is no longer hypothetical. It is being scoped.
• The streaming service becomes the label — Apple Music, Amazon Music, others follow. Two or three companies end up owning production, distribution, and listener data end to end. The major labels are negotiating from a position that has no precedent in recorded music's eras.

05. The behavioural rabbit hole

Five behavioural shifts nobody is tracking properly

This is the second tree. Five things that don't fit on the map because they need setup, and which reward the reader who's already opted in. Quotable in a pub. Slightly uncomfortable on reflection.

• The Velvet Sundown effect: harm is in the disclosure, not the listening — A million people enjoyed a fake band. They were happy until they were told. Which raises a question nobody wants to answer: is the harm in the music, or in finding out you can't tell?
  • The 97% figure is a fact about humans, not AI — If 97% of listeners can't distinguish AI from human, the interesting question isn't 'how good is AI?' It's 'when did we stop paying that much attention?' The answer is: probably before AI showed up.
• Royalty fraud farms are the actual business model for most uploaders — The 50,000-tracks-a-day figure is misleading. Most of those uploads come from a smaller number of operators running an audio-SEO play, optimising for the 1,000-stream threshold and bot-streaming their way to royalty cheques. The 'music' is incidental.
  • Spotify removing 75 million tracks isn't moderation, it's overhead — If you remove 75 million tracks and the system keeps producing more, you don't have a moderation problem. You have a category of user whose entire business is producing what you remove. The cat is not getting put back in the bag.
• The labelling system can't tell content farms from working artists — A producer who uses an AI vocal stem to fix a take gets labelled the same as a content mill cranking out lo-fi. Spotify's detection treats 'used AI' as a binary. The reputational damage falls on the legitimate user, because the content farm doesn't care about its reputation.
• Functional music dying first is a confession about what we'd turned music into — The genres being eaten first are the ones where music had already been reduced to product: sleep, study, focus, ambient. AI didn't make these genres into commodities. It just finished the job streaming started.
• Some people who couldn't afford music to use, now can — Indie filmmakers, podcasters, small-business video makers, charity ad producers. People who used to pirate library music or sit in silence now have cheap, legitimate access to something usable. Not every consequence of the flood is a loss. Some of them are, awkwardly, fine.