90,000 AI Songs a Day: Is the Black Box About to Explode?

90,000 Songs a Day That Nobody Knows Who Wrote

Deezer confirmed in July 2026 that it receives about 90,000 fully AI-generated tracks every day — over 50% of new uploads on some days in June, according to TechRadar. Before panicking at the headline, the important nuance: those tracks only generate between 1% and 3% of streams, and Deezer says up to 85% of their plays in 2025 were fraudulent and demonetized.

In other words: the catalog volume is already enormous, but the direct economic impact is still small. The worry isn't today. It's what happens when part of that catalog starts generating legitimate consumption at scale — because the rights system isn't built to ask the only question that matters: who made this song?

The Asymmetry: The Master Gets Paid, the Work Doesn't

To pay the master, the system has a simple route: recording → ISRC → distributor → uploader's bank account. Even if the authorship information is missing or false, the DSP knows who delivered the phonogram and whom to pay.

But behind that same recording sit two more layers:

The composition. Who wrote the lyrics and the music? Is there a human author at all? With what percentages? Do they have an IPI? Is the work registered? Is there an ISWC?

The performance. Who sang? Who played? Is the voice human, cloned, or fully synthetic? Which musicians are owed neighboring-rights remuneration?

If the release arrives with just a title, artist name, ISRC and audio, the master gets paid — but the composition and the performances can become impossible to identify. This already happens without AI: the US MLC holds royalties it can't match for a minimum of three years and, once identification attempts are exhausted, they can end up distributed by market share. AI brutally multiplies the number of recordings entering that circuit.

Not All of It Is Black Box: The Three Scenarios

It's worth distinguishing, because each case breaks the system differently:

1. Unregistered human work. A real composer exists, but the song circulates without authors, percentages or IPIs. This is classic black box: there's a rightsholder to pay, but the system can't identify them.

2. Fully synthetic recording. There may be no human composer or performer with rights over the result. The US Copyright Office requires sufficient human expressive contribution: writing a prompt doesn't make anyone an author. Here there's no unknown rightsholder — there's a rights vacuum: the system logs consumption of something that may have neither a protected composition nor a payable performance.

3. Invented rightsholders. The uploader claims they wrote or sang what a machine generated, or registers fictitious shares to capture royalties. That's not black box anymore: it's database contamination and registration fraud.

Neighboring Rights: Especially Tricky

In Spain, performers' remuneration for the making-available of phonograms is collectively managed by AIE. But a synthetic voice is not a performer:

If a human singer was involved, AIE needs to know who they are.

If a voice was cloned, there may be an infringement — but no new performance by that person.

If everything was generated, there may be no performer to distribute to at all.

And if the DSP or distributor can't tell these scenarios apart, the reports flowing downstream are ambiguous or plain wrong.

The full cocktail: black box + rights vacuum + identity fraud + metadata pollution.

The Claim That Actually Holds

At 90,000 songs a day, that's roughly 33 million new synthetic recordings per year on Deezer alone. No collective management organization can manually investigate who composed or performed tens of millions of phonograms. And CISAC estimates that 24% of music creators' revenues could be at risk by 2028 due to generative AI.

The conclusion isn't "AI kills music." It's more precise and more uncomfortable:

If DSPs accept the audio first and ask about the rights later, AI-generated music will make the black box grow much faster than consumption. Not just because it multiplies the songs — because it multiplies unregistered works, dubious authorships, fictitious shares, and recordings with no credited performers.

The Fix Isn't Cleaning Up Later — It's Documenting First

Trying to repair metadata years later, with the money already frozen, is the recipe that got us here. The alternative is requiring the documentation before distribution:

A declaration of human, AI-assisted or fully AI-generated content

A signed split sheet with composers, lyricists and percentages

Recording credits: who sang, who played, in what role

Real identifiers — IPI, ISNI — and society affiliations

CWR to register the composition where it belongs

That's exactly what SPLEET builds: the song is born with its splits signed by every party, each participant's identity verified with an ID document, and the metadata of all three layers — work, recording and performers — complete from minute one. We don't just register songs: we generate structured proof of who did what before the recording starts making money.

When the synthetic avalanche truly arrives, that proof will be the difference between getting paid and feeding the black box.

Document your music at the source with SPLEET