Is AI music good enough yet?

The three numbers, side by side

Each row is published by Deezer, the only large streaming service that runs and reports an AI-detection system on every upload. Dates matter here, so every figure carries one.

Read the second and third rows together. More than half of the supply, between one and three percent of the demand. Whatever is limiting AI music, it is not that people can hear the difference — the first row settles that.

How fast the supply side moved

Deezer began tagging AI uploads in January 2025. The trajectory since, from its own releases:

Nine-fold in eighteen months, on one platform, crossing the halfway line in July 2026. Deezer puts its detector at 99.8% accuracy with fewer than one false positive per 10,000 genuine tracks, so the count is a floor rather than an estimate. No other major platform publishes a comparable figure, which is the single biggest hole in this page: one company's uploads are not the whole market, and nobody else is showing their numbers.

Why the gap is not explained by quality

The obvious explanation — it still sounds wrong — is the one the evidence rules out. The Deezer/Ipsos survey put 9,000 people in eight countries (US, Canada, Brazil, UK, France, Netherlands, Germany, Japan) in front of three tracks, two AI and one human, and asked which were fully machine-made. 97% got it wrong. Of those told the answer afterwards, 71% were surprised and 52% said it made them uncomfortable; 80% said fully AI music should be labelled, and 45% said they would like to filter it out of their streaming service entirely.

That last pair is the useful part. People cannot detect it and a large minority want it removed anyway. The constraint is not audio fidelity. It is that music is chosen for reasons — who made it, what it signals, who else is listening — that survive the audio being indistinguishable.

What the labels did once it worked

The industry's own behaviour is evidence of capability, separate from any benchmark. Having sued the two leading generators in 2024, the major labels began settling and licensing once the output got good:

You do not license a technology you expect to fail. The settlements are a stronger signal about capability than any listening test, because the people signing them have the catalogues at risk.

What this bears on

This site grades whether AI capability is arriving on the predicted schedule, and separately whether arrival produces the predicted change. Music is the cleanest natural experiment available on the second question, because both sides are counted: the capability is verified by blind test, and the adoption is verified by stream share on the same platform in the same month. Almost nowhere else are both numbers published.

The answer, so far, is that they came apart. A capability can pass the human-indistinguishability bar and still move a low single-digit share of behaviour eighteen months later. Anyone forecasting labour-market or economic effects from benchmark scores alone should be able to explain why music is the exception — or accept it as the base rate. It is the same question asked on whether AI can replace knowledge workers and what it means for your job, with the rare advantage of having both halves measured.

What would change this verdict

Pre-registered, so it can be checked rather than rationalised:

• Stream share breaks 10% on any platform that publishes the number. That would turn "capability without adoption" into a transition in progress, and would be the single reading that matters.
• A second large platform publishes upload and stream shares. Today the entire quantitative picture rests on one company; a second source either corroborates it or collapses it.
• A fully AI-generated track charts on a major national chart on genuine, non-fraudulent streams. Demand, not supply, and audited.
• The fraud share falls sharply while stream share holds. That would mean the listening is becoming real rather than manufactured.

None of the four has happened as of 17 September 2026. Until one does, the honest reading is that AI music won the quality argument and has not yet won an audience.

Sources

Upload counts, stream share, fraud share and detector accuracy: Deezer Newsroom, AI music tops 50% of daily uploads (21 July 2026) and AI-generated tracks represent 44% of new uploaded music (20 April 2026). Blind-test result and listener attitudes: Deezer and Ipsos study (November 2025), 9,000 respondents across eight countries. Settlement dates: Music Business Worldwide. Where reporting gave only a relative date (“last month”), this page states the month and no day. A pending motion to expand the recordings at issue in the remaining litigation is omitted here because no primary filing could be retrieved to verify it. Deezer is the only large platform publishing these figures, so every quantitative row above rests on one company’s data.

Frequently asked questions

Yes. In a Deezer/Ipsos blind test of 9,000 listeners across eight countries published in November 2025, 97% could not correctly identify fully AI-generated tracks when played three songs, two of which were AI. Audio quality is no longer the limiting factor.

It depends on whether you count uploads or listening, and the two are wildly different. As of July 2026 more than 50% of tracks uploaded to Deezer each day — about 90,000 — are fully AI-generated, but those tracks receive only 1–3% of total streams.

The evidence rules out audio quality as the reason. In the same Deezer/Ipsos survey, 80% of respondents said fully AI-generated music should be clearly labelled and 45% said they would like to filter it out of their streaming service. Listeners appear to choose music partly for who made it, which survives the audio being indistinguishable.

Deezer reports that up to 85% of the streams generated by fully AI-generated tracks in 2025 were fraudulent. Those streams are excluded from royalty payments, so the already small 1–3% stream share overstates how much genuine listening AI music receives.

Partly, and then they licensed instead. Universal settled with Udio in October 2025 and Warner settled with Udio in November 2025 and with Suno on 25 November 2025, in each case pairing the settlement with a licensing deal for platforms planned for 2026. Sony Music has not settled, and claims from GEMA in Germany, Koda in Denmark and independent-artist class actions remain live as of September 2026.

Stream share, not upload share. This page pre-registers four triggers: AI stream share passing 10% on any platform that publishes it, a second large platform publishing comparable figures, a fully AI track charting nationally on genuine streams, or the fraud share falling sharply while stream share holds. None had happened as of 17 September 2026.