AI-Generated Music Saturation: What Deezer's Numbers Do and Don't Tell You

The advice circulating in board decks and investor memos right now is tidy enough to fit on one slide: AI-generated music saturation is a volume problem, not a revenue problem. Tens of thousands of machine-made tracks land in streaming ingestion pipes every day, almost nobody plays them, and because streaming pays out of a pro-rata pool weighted by plays, tracks with no plays take no money. File under noise. Monitor it, don't budget for it.

That advice is roughly right in the one place people bother to test it. It falls apart in four places they don't.

I say this as someone who spends most weeks on the production side of this rather than the analysis side — scoring cues, cutting loops, arguing with a client about whether the bed under a trailer needs a live cello. The reason the slide bothers me is not that it is wrong about the arithmetic. The arithmetic is fine. It is wrong about where the cost lands, and it is built on a number that almost nobody who quotes it has looked at closely.

Where the advice holds up

Start with the part that survives scrutiny, because it genuinely does.

Under a pro-rata model, the royalty pool for a given territory and period is divided according to each recording's share of total qualifying streams. A track with zero streams has zero share and receives zero. Adding a hundred thousand zero-stream tracks to a catalog does not move a single unit of currency out of anyone's account. This is arithmetic, not opinion, and it does not change under a user-centric model either — a subscriber's fee is allocated across what that subscriber actually played, so unplayed inventory is still invisible.

The major platforms have also, as of writing, layered policy on top of the arithmetic. Minimum-stream thresholds — the requirement that a recording clear some floor of annual plays before it earns anything at all — were introduced partly to stop micro-payout administration from eating more than the payouts were worth, and they have the side effect of structurally excluding the long tail of never-played uploads from the pool. Deezer has said publicly that it excludes tracks its detection system flags as fully AI-generated from algorithmic recommendations and editorial playlists, and that it does not pay out on streams it identifies as fraudulent. Spotify has described adopting industry metadata standards for AI-disclosure fields alongside expanded spam filtering. Whatever else you think about those measures, they are the platforms doing exactly what the tidy slide assumes they will do.

Storage and delivery costs per additional track are small enough to round off. And the top of the catalog is not being displaced: the recordings that generate the overwhelming share of consumption are the same recordings that generated it before, and nothing in the public data suggests a machine-made track has taken a top-tier slot on merit rather than on manipulation.

So: if your only question is whether a flood of unplayed AI uploads mechanically transfers money from human artists to prompt operators through the pro-rata split, the answer is no, and the people telling you not to panic are correct.

That is a narrower question than the one most executives think they are asking.

The headline figure is a classifier output, not a census

Every saturation number in circulation traces back to a small number of platform disclosures, and the most-cited of them come from Deezer, which built an AI-detection system, has been publishing its findings since early 2025, and has revised the figure upward at nearly every update — from a share of daily uploads in the low tens of thousands to figures several times that within roughly a year. Deezer deserves credit here. It is close to the only platform putting a number in public at all, and the industry's habit of treating its transparency as a free data feed while extrapolating wildly from it is the industry's problem, not Deezer's.

But understand what the number is. It is the output of a classifier, with a decision threshold, running on one platform's ingest stream. Deezer has published the headline share. It has not, to my knowledge, published a confusion matrix: the false-positive rate, the false-negative rate, or how those rates shift across genres. That matters more than it sounds. A detector tuned for the sonic fingerprints of the two or three largest generation models will flag their outputs efficiently and will be blind, by construction, to outputs from a model it has not seen. It will also, sooner or later, flag a bedroom producer working entirely in a DAW whose mixdown happens to sit in the same statistical neighborhood.

Then there is the definitional problem, which is worse. "AI-generated" is being applied as a binary label to something that is a gradient. A record cut this year might use AI-assisted mastering, machine stem separation on a sampled break, algorithmic vocal tuning, a generated pad buried at −24 dB under a live bass, and a human writing every note that matters. Detectors are trained to catch fully generated output, so the current numbers mostly describe that end of the spectrum. As hybrid workflows become the default in professional production — and they are becoming the default — the boundary the detector is drawing will stop corresponding to anything a rights-holder, a licensor, or a regulator actually cares about.

The practical consequence for anyone building a model off these figures: treat the number as an index, not a measurement. The direction of travel is real and steep. The level is soft, and the level is what gets quoted.

Do AI-generated tracks actually get streamed?

Yes — and a large share of those streams are not people. Deezer has said that the majority of streams going to fully AI-generated tracks on its platform were fraudulent, meaning automated traffic aimed at the payout pool rather than listeners choosing to press play. So the comforting claim that "nobody listens to it" is half true in a way that inverts its own conclusion. The tracks are not competing for attention. They are competing for the pool directly, and skipping the attention step entirely.

This is the reframing that should be doing the work in your risk model. Streaming fraud is not new, and it was never limited by imagination. It was limited by supply. Running a stream farm requires plausible-sounding assets in volume — enough distinct recordings that the traffic pattern does not resolve into an obvious cluster, each one credible enough to survive a spot check. Historically, acquiring those assets cost something: buying out library cues, commissioning throwaway production, or ripping and mangling existing recordings in ways that risked content-matching.

Generation models took that cost to roughly zero and removed the risk of a fingerprint match at the same time. What looks like a creative-industries story is, in the segment that matters financially, a fraud-economics story wearing borrowed clothes. The constraint on manipulation loosened, and the volume that followed is the market clearing at the new constraint.

That has a cost, and it is a real one, but it does not appear where the tidy slide looks for it. It appears in detection infrastructure, in clawback and adjustment processes, in the reconciliation burden between platforms and distributors, and in the distributor relationships that get terminated. It is an operating expense line and a trust-and-safety headcount line. Nobody in the pro-rata debate is watching either.

The cost is per-track, not per-stream

Here is the structural point the royalty-dilution framing misses entirely: most of the cost of a recording existing is incurred before anyone decides whether to play it.

Ingestion and transcoding. Metadata normalization and dedupe. Identifier allocation and the downstream mess when identifiers are allocated badly. Rights registration and the matching process at collective management organizations, where a work with thin or synthetic credit data does not match cleanly and drops into a manual queue. Dispute and takedown handling. Editorial and catalog operations. Legal review when a flagged upload turns out to have a real complainant behind it.

Every one of those scales with track count. None of them scales with stream count. A zero-stream upload is not a zero-cost upload; it is a track that consumed the full unit cost of existing and returned nothing against it. Publishing and CMO back offices in particular are staffed and budgeted against a historical rate of new works entering the system. That rate is the assumption under stress, not the payout formula.

The four denominators

Most of the confusion in executive conversations about this comes from four different measurements being quoted interchangeably. They are not interchangeable, and the gap between the first and the last is enormous.

The headline saturation figures are upload-share numbers. They are being cited in rooms where the question on the table is revenue exposure. Those are the two ends of the table, separated by three filters — acceptance, playback, and payment eligibility — each of which removes a large fraction. If a slide moves from an upload percentage to a conclusion about royalties without naming those filters, the slide is not analysis.

Where the saturation actually concentrates

Aggregate share hides distribution, and the distribution is the whole story.

Machine generation is competent, right now, at a specific and commercially significant thing: undemanding instrumental inventory. Ninety seconds of warm ambient at 70 BPM in D minor. A neutral corporate underscore that resolves cleanly at 0:30 and 1:00. Lo-fi with tape wobble and a filtered break. Sleep, focus, spa, retail bed, background. No vocal, so the hardest problem is sidestepped. No arrangement risk, because the brief is to not be noticed. It sits fine at −14 LUFS and it will not embarrass anyone.

That is also, precisely, the inventory that production-music libraries and functional playlists have sold for decades. It is the most commoditized tier of recorded music, the tier where the buyer was never choosing a specific artist, and therefore the tier with the least attachment protecting it.

So the exposure is not "AI takes the top of the charts." It is price compression in the commodity tier. I see the shape of it from the production side: the scoring work where a director wants a specific point of view has not moved, and the conversations about it have not changed. The library cue budget is where it shows up. A client who used to license three tension beds for a corporate piece now generates a dozen candidates, keeps two, and pays a subscription instead of a sync fee. The money did not vanish; it changed which line item it comes out of and how much of it there is.

For a publisher or a library owner, that is a revenue-mix question worth modeling seriously, and it is invisible in any analysis that reasons about the catalog in aggregate. For a frontline label, it is close to irrelevant. Both of those firms are currently reading the same headline number and drawing the same conclusion from it, which means at least one of them is wrong.

The generation side has its own risk surface

One more thing that gets left out of the saturation frame, because it lives on the other side of the transaction.

When a user-data breach at a major generation platform surfaced through breach-notification channels, the industry read it as a consumer-privacy story. For anyone running a studio, a post house, or a label with production staff, it is a vendor-exposure story. Generation platforms hold account records, and they hold prompts. Prompt text on a working project routinely contains unreleased material: artist names, project titles, release timing, the brief a client sent under NDA. That is a data-classification question your team has probably never been asked, about a vendor that probably entered the building on a personal credit card.

The litigation picture moved in the same period. The major-label actions against the largest generators converted, as of writing, into licensing arrangements rather than judgments, which resolved less than the headlines suggested. It moved the fight from whether training on catalog is permissible to what a licensed pipeline has to document, and that makes provenance metadata the contract surface for everything downstream. Whichever tool your team uses — City of Punk's own included — the question that survives an audit is whether you can produce a generation record for a given asset, with a date, a model, and a license state attached. Most shops today cannot, and are discovering this at the moment a delivery spec asks for it.

Before a saturation figure goes in the deck

Eight questions. If the person presenting cannot answer at least six, the number is decoration.

• Which denominator? Uploads, catalog, streams, or payable revenue. Name it explicitly.
• Detector or disclosure? A classifier's guess and a distributor's declared flag are different data with different failure modes.
• What threshold, and what error rates? If the false-positive rate is unpublished, say so in the footnote rather than in the appendix nobody reads.
• Submitted or accepted? Uploads rejected at ingest never became catalog.
• Gross or post-fraud-filter streams? For this category, the gap between the two is the finding.
• What share of payout-eligible streams? After minimum thresholds, this is the only figure with money attached.
• Which segments? Break it out by functional versus vocal-led inventory, or the average will tell you nothing about your own mix.
• Whose number is it? Every platform publishing detection data is also selling the credibility of its detection. That does not make the data wrong; it makes the incentive worth stating.

The more honest version of the rule

The tidy slide was trying to say something true: this is not a royalty-dilution event. Keep that. Then extend it, because the same facts that make it not a dilution event make it three other things.

It is a unit-cost event, because the expense of a track existing is incurred at ingestion and registration rather than at playback, and those functions are budgeted against a submission rate that no longer holds. It is a fraud-supply event, because the binding constraint on stream manipulation was the cost of plausible assets and that constraint is gone. It is a price-compression event in the commodity tier, concentrated in functional and library inventory, invisible in any catalog-wide average. And it is a measurement problem, because the numbers everyone is quoting describe uploads on one platform through one classifier, and are being used to reason about revenue across the market.

None of that is cause for alarm. All of it is cause for putting the exposure on the right line. The firms that get this wrong will not be the ones that panicked; they will be the ones that filed a real operating-cost and revenue-mix problem under "industry noise" because the pro-rata math checked out.

The myth is that AI-generated music saturation does not matter because nobody listens to the output. The more accurate version is that almost nobody listens to the output, and that is exactly why the cost lands in ingestion, fraud detection, and the commodity tier of the catalog instead of the royalty pool everyone is watching.