Half of New Uploads Are AI. Almost None of the Streams Are.

Deezer put out two numbers on July 21, 2026, using June data, and they tell very different stories. The first: fully AI-generated tracks passed 50 percent of daily new-music uploads in June 2026, nearly 90,000 tracks a day. The second, from the same report: those tracks account for between 1 and 3 percent of total streams. One measures what gets pushed into the pipe. The other measures what anybody actually listens to. Almost every headline quotes the first and stops, which leaves working artists with the impression that half their competition is now a machine. It isn't.

The two numbers everyone runs together

Deezer has been publishing this since it launched AI detection in January 2025, when the figure was about 10,000 tracks a day. By April 2026 it was 44 percent of uploads, around 75,000 a day. Two months later it crossed half.

But upload share measures supply, and supply stopped being the constraint long before AI arrived. Luminate counted roughly 106,000 new tracks delivered to streaming services a day across the industry in 2025, and most of the catalog already sat at near-zero streams. Deezer's own intake runs well above that. Pouring more tracks into a system where most human-made music already goes unheard doesn't change the competitive picture much. The constraint was never how many songs exist. It was always attention.

The listening number says it directly: one to three percent. If AI music were quietly eating the audience, that figure would be climbing toward the upload figure. It isn't.

Why the flood barely reaches listeners

The gap is enforcement, not accident. Deezer says it excludes AI-detected tracks from algorithmic recommendations and editorial playlists, which cuts off the main road any track travels to reach a stranger. A song nobody is recommended and nobody searches for gets played by nobody, however many of them you upload.

Spotify came at it from a different angle and landed somewhere similar. In September 2025 it announced it had removed over 75 million spammy tracks in the previous twelve months. It also added impersonation enforcement, AI disclosure credits, and a music spam filter aimed at "mass uploads, duplicates, SEO hacks, artificially short track abuse, and other forms of slop." Spotify did not ban AI music. Its stated position is that AI is fine as a tool and the target is content farms and bad actors.

There's a fraud dimension too. Deezer reports that up to 85 percent of the streams that fully AI-generated tracks did produce in 2025 were fraudulent. So a good chunk of that 1 to 3 percent is bots draining the royalty pool rather than anybody listening. That makes real human demand for this material even smaller than the 1 to 3 percent suggests.

Velvet Sundown is the clearest case. The AI-generated "band" crossed a million Spotify monthly listeners in early July 2025 and peaked near 1.4 million on the back of enormous press coverage. By mid-August it was down to roughly 594,000 (Music Ally). Press coverage produced the spike, and the listeners left when the coverage did.

Where the slop actually costs you

The flood still does damage, just not where artists expect.

Discovery surfaces get noisier. Every recommendation system is trying to infer taste from behavior. It now does that inside a catalog where a growing share of items are made to game the system rather than to be enjoyed. Platforms respond by getting more conservative and leaning harder on signals that are expensive to fake. That's rational, and it quietly raises the bar for a real artist with a small footprint.

Ad targeting degrades the same way, for the same reason. The behavioral pools that Meta and TikTok build audiences from get muddier when a slice of the underlying activity is synthetic.

The third cost is attention from the people who decide things. Editors, playlist curators, and A&R staff are all now filtering harder because the volume forces them to.

Clean signal is the scarce thing now

When machine-made supply is unlimited and cheap, the scarce commodity is proof that real people chose your song on purpose. That's measurable.

Save rate is the test. It's the share of listeners who add a track to their own library, and it's expensive to fake and rarely happens by accident. Below 8 percent, something's off. Between 8 and 15 percent is solid, above 15 percent is strong, and above 20 percent is exceptional. Bought and bot-driven streams sit near zero, which is why they're detectable. We've written separately about why save rate beats stream count as the number to run a campaign on.

Real ad-driven listening produces the opposite profile. Across our campaigns a click converts to roughly 0.38 saves in Rap and HipHop, 0.51 in Pop, and 0.88 in Rock and Alternative, varying with genre and creative quality. Those are people who heard a song, liked it enough to keep it, and left a trail that reads unmistakably human. That trail is what feeds the recommendation systems we broke down separately, and it's what compounds into algorithmic lift over a catalog's life.

So the strategic read on 50 percent is almost the opposite of the panic. The flood is real, it's in the upload pipe, and it's making legible human listening more valuable, not less. Artists chasing volume will struggle, because a machine can produce volume for nothing. Artists who can point at a save rate and show people chose the song will do fine. We build campaigns to produce that number, judged on the metrics that separate listening from traffic. You can see it at scale in a six million stream Spotify playlist campaign we ran on Meta.