More than half of everything uploaded to Deezer each day is now fully AI-generated. Those tracks pull 1 to 3 percent of streams, and up to 85 percent of even that was fraud.
Deezer’s newsroom reported on July 21, 2026 that fully AI-generated music hit nearly 90,000 tracks delivered per day at its June peak, crossing 50 percent of all daily uploads. The same reporting puts AI tracks at 1 to 3 percent of total streams. Supply crossed the halfway line and demand did not move. That gap is the whole story.
The 85 percent figure is the one I keep returning to. If AI tracks take 1 to 3 percent of streams, and up to 85 percent of those streams were artificial and got demonetized, then the genuine human listening attributable to the largest content category by upload volume rounds to somewhere near nothing. Deezer did not publish a clean post-demonetization stream share, so I will not pretend to compute one. The direction is still clear. The flood is competing for payout slots in a royalty pool rather than for listeners, and the platform caught it.
From 39% to over 50% of daily uploads in six months
The growth curve is well documented, because Deezer published it three times in the same year. In January 2026, about 60,000 fully AI-generated tracks arrived per day, roughly 39 percent of daily uploads. By April, almost 75,000 per day, roughly 44 percent. By June, nearly 90,000 per day and over 50 percent. That is an 11-point share gain in six months while the absolute daily volume rose by half. On accumulated stock rather than flow, Deezer detected more than 13.4 million AI-generated tracks across all of 2025.
Deezer is one platform, and a mid-sized one. The broader check comes from SubmitHub’s SH Labs detector, which analyzed over a million tracks released globally in July 2026 and found AI involvement in 38.5 percent of them: 23.2 percent fully AI-generated, 15.3 percent AI-generated audio that a human then modified or processed. Euronews reported those figures on August 20, 2026.
Note the shape of that split. The hybrid category, human-edited AI audio, is not small at 15.3 percent, about two thirds the size of the fully synthetic category. That is the category that breaks every simple policy, and I will come back to it.
Why AI tracks get uploaded when almost nobody streams them
Because upload is nearly free and the payout mechanism pays per stream from a shared pool. If you can generate 90,000 tracks a day at negligible marginal cost and then inflate streams on them, you do not need anyone to like the music. You need the accounting to clear. Deezer’s January statement that up to 85 percent of streams on fully AI tracks were fraudulent or artificially inflated, and that those streams were demonetized, says plainly what the economic engine was.
What the numbers do not tell you matters as much. The 1 to 3 percent stream share and the 85 percent fraud rate both come from Deezer’s own detection and its own catalogue. I have no independent audit of either figure, and nobody has published a reconciliation between the 85 percent fraud number from January and the 1 to 3 percent stream share from April and July. Directionally strong, precisely uncertain.
The detection side has moved faster than I expected. SH Labs is reported to have hit over 99 percent accuracy on confidential industry benchmarks, with later coverage citing 99.4 percent in third-party testing. Deezer shipped a free consumer-facing detector for playlists on June 11, 2026, and its detection system won a Billboard award announced September 16, 2026. Those are production systems running against a live firehose, not research prototypes.
Then SubmitHub did the thing most platforms will not do. On September 17, 2026, Digital Music News reported that SubmitHub, founded by Jason Grishkoff, banned AI music from its promotion services outright, blocking any track scored 85 percent or more likely to contain AI audio. Roughly 11 percent of tracks uploaded in the preceding month crossed that threshold. Secondary coverage estimated the ban could cost around 10 percent of revenue. Writing your own lyrics buys no exemption: the policy covers hybrid human-edited AI tracks as well as fully synthetic ones.
**A company voluntarily gave up an estimated tenth of its revenue to enforce a line, and drew that line to include human-edited output.**
Detection economics, not music, is the transferable lesson
I read this as a live test of a pattern that other content platforms are about to run. Three findings transfer.
First, volume and value decouple violently. Going from 39 to over 50 percent of upload share changed the composition of the catalogue and did essentially nothing to listening behaviour. If your input is cheap and your audience is finite, expect the input to saturate long before it captures attention. Plan capacity for the flood; do not plan revenue for it.
Second, fraud is the actual product. The 85 percent number reframes the whole phenomenon. The flood was mostly a payout-extraction scheme wearing songs as a costume. My read: when you see a sudden synthetic content surge on any platform with a revenue-sharing mechanism, your first hypothesis should be fraud against the mechanism rather than a new creative wave.
Third, detection is now good enough to act on, and the hard part has moved to policy. A 99 percent detector is useless if you have not decided what a 90 percent score means. SubmitHub chose 85 percent as its blocking threshold and accepted the false positives that come with it. That is a business decision dressed as a technical one, and it is the decision most organisations are currently avoiding.
The hybrid category is where this gets genuinely hard. SH Labs found 15.3 percent of July’s global releases were AI audio touched by a human. Detection can flag the audio; it cannot measure how much a person contributed. SubmitHub resolved that by refusing to draw a gradient at all: 85 percent AI-likely means blocked, lyrics you wrote yourself included. I think blunt thresholds are the right call for now, because a graduated policy needs a measure of human contribution that nobody has. I also think that call will age badly within two years, once AI-assisted editing is standard in normal production workflows and a flat ban starts catching ordinary work.
There is a provenance answer to this problem, in theory. Content Credentials would let a track carry a signed record of how it was made, removing the need to infer origin from the waveform. In practice, as I covered in the C2PA breakdown, most platforms strip that metadata on upload, which is why we are stuck running statistical detectors against finished audio instead of reading a manifest. Detection is the fallback for a provenance system that did not get deployed.
Upload share is a supply metric being read as a market metric
Half of Deezer’s daily uploads being synthetic tells you the cost of producing a track fell to roughly zero. It tells you nothing about what listeners chose, and the 1 to 3 percent number says they mostly did not choose it.
I would not read that 1 to 3 percent as comfort for working musicians, though. The pressure on artist income shows up through licensing, sync, session work, and library music rather than through AI tracks beating human ones head-to-head on the same playlist, which is why artists report lost income even while winning in court. A low stream share for synthetic tracks and a shrinking income base for humans are entirely compatible.
For anyone running a platform with user-generated content and a payout mechanism, the sequence Deezer and SubmitHub have demonstrated is the one to copy: measure the inflow, separate genuine engagement from inflated engagement, demonetize the inflated portion, publish the numbers. Deezer published its share three times in six months and let everyone watch the curve. **That transparency is why we can have this conversation with figures instead of vibes.**
The number I want next is the post-demonetization stream share for AI tracks, broken out from the 1 to 3 percent. Nobody has published it. Until someone does, we know the flood is real and we know the audience is not, but we do not know the exact size of the hole in the middle.
