How much music is actually AI? Qobuz and SIQA's numbers, read carefully

Two sets of numbers landed in September that, read together, answer a question the industry keeps getting wrong: how much music is AI? On 24 Sep 2026, Qobuz said only 0.38% of its streams come from tracks it has identified as AI-generated. Earlier the same month, the Sonic Intelligence Academy (SIQA) published a mid-year report saying fully AI-generated work is shrinking among the creators it charts. Neither number means what a headline would suggest. Both point the same way.

What did Qobuz announce?

The Paris-based streaming service now shows a tag in its app on music it has flagged as AI-generated, and keeps that music out of its editorial selections, playlists and featured slots. The detection system was built in-house; Qobuz doesn't describe how it works. Alongside the 0.38% figure, Qobuz says 60% of streams on AI-identified tracks are currently judged fraudulent by its anti-fraud systems and excluded from royalty payouts. It also says it removed more than a third of AI-identified albums with no listening activity from its search engine. Zinstrel reports that those releases still turn up on an exact-title search, so they were de-indexed rather than deleted.

Deputy CEO Georges Fornay framed it as a labelling job: the task is no longer just detecting AI music but letting listeners see it clearly.

So how much real listening is AI?

Less than 0.38%. Zinstrel applies the 60% fraud share to the 0.38% and gets roughly 0.15% of all Qobuz streams from people actually listening. That is Zinstrel's arithmetic, not a Qobuz figure, and Zinstrel itself notes Qobuz hasn't confirmed both numbers use the same base. Qobuz also gave no measurement window and no total stream volume.

Deezer's figures land in the same place. Its January release put fully AI tracks at 1-3% of streams, with up to 85% of those streams fraudulent in 2025 depending on the month. Net out the fraud and the real share is a fraction of a percent. Meanwhile supply keeps climbing: in July Deezer said it receives nearly 90,000 fully AI-generated tracks a day, more than 50% of daily deliveries on peak days.

One caveat cuts in both directions. Both services keep flagged music out of their own recommendations, so low listening partly reflects their policy, not only listener taste.

What does SIQA's report actually measure?

SIQA runs a chart for AI-made music. Its mid-year report covers 1 Apr to 20 Aug 2026: 1,743 submissions from 886 artists in 54 countries. The headline finding on its public page is that 86% of submissions involve meaningful human creative contribution. The full tier breakdown sits behind a sign-up, so we're relying on AIMusicpreneur's write-up (15 Sep) for it: AI-assisted rose from 48.4% in Q1 to 51.7%, a "human + AI hybrid" tier (artists using a voice model built only from their own voice) rose from 32.4% to 34.4%, and fully AI-generated fell from 19.2% to 13.9%.

Read that carefully. This is a self-selected sample: creators who chose to submit to an AI chart and picked their own tier, which SIQA then verifies. SIQA describes it as creator-disclosed data. It tells you about people who make AI music on purpose and want credit for it, not about the whole catalogue, and certainly not about the roughly 90,000 fully AI tracks a day that Deezer is filtering. AIMusicpreneur also points out that declaring "fully AI" got costlier over the same months, with chart bodies moving to exclude it, so some of the shift may be people labelling more carefully rather than working differently.

What artists are actually doing

The evidence from musicians fits the hybrid picture. One artist on X points to the move from fully AI submissions towards AI-assisted ones as the trend worth noticing. A songwriter describes writing every lyric themselves and using AI only to sing them. That is a typical hybrid case: the authorship, the words and the decisions are human, and one part of the performance is generated.

For artists like that, the stakes are rights and income. A tag designed to catch fully generated content can land on a track where a human wrote everything. Qobuz's tag keeps flagged music out of editorial placements, and its fraud filter strips royalties from flagged streams. Zinstrel notes Qobuz's charter accepts AI in mixing, mastering and composition and concedes detection is imperfect where humans were partly involved, and it sends artists who think they were wrongly tagged to their distributor. That is a long way round for an independent musician trying to find out why a release vanished from playlists.

The practical advice is dull but useful: keep your lyric drafts, session files and, if you use a voice model, proof that it is your own voice. AIMusicpreneur makes the same point, and it is the only defence an artist fully controls.

What we think

The numbers describe three different populations, and policy keeps treating them as one. There's a flood of fully generated uploads that almost nobody hears, much of whose listening is fraud. There's a small, visible group of people who make AI music deliberately. And there's a growing middle of human artists for whom AI is one tool in the session.

Platforms have mostly solved for the first group, and the solution works: demonetise fraud, keep flagged tracks out of recommendations, and the real listening share drops to a fraction of a percent. The second-order risk is that the same blunt instruments now apply to the third group. A binary "AI" tag can't tell a songwriter with a generated vocal from a farm uploading thousands of tracks a day, and the artist is the one who pays when it guesses wrong: lost placement, withheld royalties, an appeal that runs through a distributor.

Who wins under the current rules? Services that can say they're cleaning up, and fraud filters that work on the slop. Who loses? Human-led hybrid creators, who are probably the fastest-growing group and the least able to argue with a classifier. The fix is labelling that distinguishes assist from replace, ideally with the artist declaring which parts used AI, plus a human-reviewed appeal at the platform, not the distributor. Tools are already moving that way, with AI working inside a human-led session, such as the co-producer in Veena's browser DAW. Policy should catch up with the people who use them.