AI is now the majority of Deezer's daily uploads. Retroactive takedowns are starting.

On July 21, 2026, Deezer disclosed that fully AI-generated music crossed 50% of all new uploads for the first time in June, averaging around 90,000 AI tracks per day. In the same statement, Deezer announced it will now take down AI content used to inflate streams fraudulently, plus AI tracks with no listener activity in the last six months. This is the second DSP-side enforcement event in eight days, after Qobuz/FUGA started rejecting AIGC at delivery on July 13. Here's the shape of the pincer that's forming, and what it changes for distributors and human artists.

The numbers

Deezer has been publishing AI-upload volume since late 2024. The trend line, in the two most recent data points:

Two months, +20% volume, +6 percentage points of share. If the curve continues, AI will be two-thirds of Deezer's new uploads before the end of the year. This is the first time a mid-tier DSP has publicly confirmed that the majority of everything hitting its ingest queue is synthetic.

44% was already the reason distributors started thinking about pre-upload filters. 50% is the reason DSPs start doing catalog cleanup.

What Deezer said it will do

Two categories of takedown, described in the same disclosure:

1. AI content used for stream fraud. Tracks whose listen patterns match the classic bot-inflation shape: high volume from low-diversity sources, bursts around payout thresholds, disproportionate skip rates. This is anti-fraud enforcement — an extension of what Deezer already does against click farms — that now uses AI-provenance as a co-signal. If the track is AI and the streams look fake, both go together.
2. AI content with zero listener activity for six months. This is the more unusual half. It's catalog hygiene: Deezer is removing AI tracks that never mattered anyway. The economic logic is simple — indexing, storage and rights admin cost real money per track, and a track with no listeners for six months has negative unit economics. If it's also AI, there's no artist relationship to preserve, so the removal is cheap.

The second bucket is the one that will surprise catalog owners. It applies retroactively. A track that shipped clean six months ago and never found an audience is now a takedown candidate purely on catalog-cost grounds. This is a structurally different failure mode from Qobuz/FUGA rejection at delivery — that one blocks new tracks, this one prunes existing ones.

The pincer: entry filter plus catalog cleanup

Eight days ago, Qobuz turned on a distributor-tier filter that blocks AIGC before delivery. Today, Deezer turned on a DSP-tier cleanup that removes AIGC after delivery. Together they define both edges of the pipe:

Any AI track has two ways to get killed now: it can be blocked on the way in, or removed after the fact. Neither of these existed in production form a month ago. The prediction we made in the Qobuz post — "other DSPs follow within 6-12 months" — collapsed to eight days.

The forecast now flattens into a much simpler statement: every major DSP will have some form of AI enforcement live within the next quarter. Deezer's move telegraphs the shape it will take at Spotify, Apple and Amazon: retroactive removal tied to stream-fraud and unplayed-catalog signals. Nobody has to build a new detection stack — they already run AI detection internally for royalty-share purposes. All they have to do is start acting on the output.

The new failure mode: false-positive takedowns of human tracks

Here is the part that changes the audience for AI detection. Until this week, DistroShield's core buyer was the distributor — the actor in the chain that needed to filter at ingest to avoid downstream refusals. Retroactive DSP takedowns add a second audience: the human artist whose track gets removed by mistake.

Retroactive takedowns run against every DSP's internal AI classifier at scale. At Deezer's stated volume, even a 99% accurate classifier produces hundreds of false positives per day. When Deezer takes down a track that mostly likely wasn't listened to much, the artist's remedy is an appeal — and the appeal only works if the artist can produce evidence that the track is human, not just the artist's word for it.

A signed audit certificate from an independent detector is exactly that evidence. Model version, per-class probabilities, timestamp, cryptographic hash, no relationship to either the DSP or the distributor. Something a human artist attaches to a takedown appeal and something a distributor attaches to a delivery escalation. The same artifact serves both sides.

What to do this week

Two different audiences, two different actions.

If you're a distributor:

1. Do everything from the Qobuz/FUGA piece — pre-upload check, distinguish 100% AI from AI-assisted, keep the audit trail. That's now table stakes.
2. Add one thing on top: for every human track you deliver, capture and store the DistroShield certificate at ingest. Six months from now, when a DSP retroactively removes a human track because it looked suspicious on second pass, your artist will call you before they call the DSP. Having the certificate on file lets you answer with evidence instead of promises.

If you're a human artist:

1. Run each of your releases through a detector once and save the certificate. Not because you think your track is AI — because a DSP might, and evidence in advance is worth much more than evidence after a takedown notice.
2. If you find a track of yours has been misclassified — for example if DistroShield's free lookup returns uncertain or hybrid on a track you know is human — pay for the full report to see which signals fired. Sometimes the fix is in the mix or in the metadata, and correcting it upstream is cheaper than fighting a takedown downstream.

What the DistroShield certificate contains

A single JSON payload plus a signed PDF. The payload is what you or your distributor forward when appealing a takedown or escalating a delivery. Concretely:

{
  "track": {
    "title": "...",
    "artist": "...",
    "isrc": "..."
  },
  "verdict": {
    "classification": "human",
    "ai_score": 0.02,
    "confidence": 0.96,
    "recommendation": "deliver"
  },
  "model": {
    "primary_version": "v9-4class-2026-06-23",
    "attribution_version": "v8-attribution-2026-05-04",
    "primary_class_probs": {
      "human": 0.96,
      "hybrid": 0.03,
      "unknown_ai": 0.01,
      "suno_unlicensed": 0.00
    }
  },
  "signals": {
    "fingerprint_match": null,
    "cross_distributor_matches": [],
    "metadata_validation": { "score": 0.97, "issues": [] }
  },
  "analyzed_at": "2026-07-21T18:30:00Z",
  "certificate_hash": "3f7c9e2b1a4d..."
}

The certificate_hash is the SHA-256 of the signed PDF version of the report. It's what makes the artifact tamper-evident: if a DSP later disputes the certificate, the hash on file at DistroShield either matches or it doesn't. The signed PDF version is what the artist or distributor emails to appeals@dsp or support@distributor. Same document, two audiences.

The window

Two weeks ago the industry conversation was still "should DSPs act on AI provenance." Today the conversation is "which DSP acts next, and how much of the human catalog is caught in the sweep." That is a fast collapse.

The distributors that had pre-upload filters wired in July 13 are already halfway to the certificate layer — they have the API integration, they just have to persist the response per delivery. The distributors that didn't have a filter on July 13 now have two things to build simultaneously, on a compressed timeline, against a policy surface that is moving faster than their engineering roadmap. And the artists sitting between them and the DSPs are the ones with the least tolerance for false positives.

DistroShield's public lookup returns a verdict on any track, free. If the verdict is human, a signed certificate costs $5 and takes about a minute. That certificate is what you attach to a takedown appeal in six months when a DSP's classifier gets it wrong.