A distributor queue fills with ten thousand new tracks before lunch. Almost nobody presses play. That gap — enormous supply, negligible demand — is the thing the AI music industry keeps mistaking for a breakthrough, and it is the one gap the platforms themselves have started to measure.
Key takeaways
- Deezer reports that about 75,000 fully machine-made tracks arrive every day, roughly 44% of all daily uploads as of April 2026, with a June peak above half of everything delivered.
- The same company says those tracks account for only 1–3% of streams, and that up to 85% of their 2025 streams were fraudulent and therefore demonetised.
- Spotify began labelling artist profiles judged to be synthetic personas in September 2026 and excludes them from algorithmic and editorial recommendations.
- China takes a different route: a four-ministry labelling rule that took effect on 1 September 2025 makes both visible labels and hidden metadata a legal obligation.
Supply is solved. Demand is not.
Deezer is the only major streaming service that publishes detection numbers rather than waiting to be told, and its disclosure is stark. As of April 2026, about 75,000 fully machine-made tracks were uploaded each day, representing roughly 44% of total daily uploads. By the June peak the figure passed 90,000 a day and more than half of new deliveries. In 2025 alone the platform tagged 13.4 million such tracks.
Then the second number arrives and changes the story: fully machine-made music accounts for 1–3% of plays. Uploads are a flood; listening is a trickle. The tracks that do reach listeners tend to arrive through background and mood playlists, autoplay and generic search terms rather than through anything a listener chose.
The economics follow. Deezer says up to 85% of streams on fully machine-made tracks in 2025 were fraudulent, and those streams were removed from royalty calculations. Its response has three parts: tag the track, exclude it from algorithmic and editorial recommendation, and strip fake plays out of the royalty pool. A Deezer–Ipsos study the company cites found 80% of listeners believe fully machine-made music should be clearly labelled.
Spotify's answer is a badge, not a ban
Spotify approached the same problem from the identity end. From September 2026, artist profiles that appear to be synthetic carry an "AI persona" badge — either self-declared by the artist or applied after review — shown on the profile, in search results and next to tracks in playlists. Music from those profiles is kept out of editorial and personalised recommendation unless a listener already follows the artist.
The design is deliberate: nothing is banned, but the algorithmic oxygen that turned anonymous uploads into playlist fixtures is cut off. Spotify had already spent 2025 on the plumbing — a spam filter aimed at mass uploads, duplicates and very short tracks, a rule that voice clones of a real artist require that artist's permission, and support for an industry credit standard that lets human songwriters disclose how a track was made. The company said it removed more than 75 million spammy tracks in the year before that policy.
China: labelling as a legal duty
China's route is regulatory rather than product-led. the Measures for Labelling Synthetic Content Made by AI (人工智能生成合成内容标识办法), issued by the Cyberspace Administration of China together with the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration, took effect on 1 September 2025, alongside a mandatory national standard (GB 45438-2025).
The structure is worth understanding because it differs from a badge:
- Explicit labels must be perceptible to the user — text, sound or graphic marks; for text content, wording such as "AI生成" at the beginning, end or another prominent position, in legible type.
- Implicit labels are written into the file's metadata, with field names required to carry the AIGC marker, and covering generation status, the generating service provider, the distribution platform, a unique identifier and a digital signature or hash.
- Removing, altering, forging or hiding those labels is prohibited, as is supplying tools to do so.
Platforms have layered their own detection on top. Tencent Music has said in its ESG reporting that it uses audio-feature models to identify and tag content with machine-made characteristics; NetEase Cloud Music asks uploaders to declare synthetic content themselves and then audits it.
What this means if you make music
According to Li An, chief scientist at BrainNet (脑机网), the distribution layer is where the value now sits: making a track is close to free, so what is scarce — and therefore paid for — is placement, playlist trust and the ability to prove a human was involved.
Practically, the workflow has split in two. On one side are producers who use models to sketch demos and then pay arrangers, singers and mixers to finish them; the scarce skill becomes judgement about which take to keep. On the other are operators pushing volume into a royalty pool that is being actively defended by fraud filters. The second route is where the 1–3% number bites.
Honest limitations
The percentages in this piece come mostly from Deezer, which publishes them because detection is its own product and its commercial interest is in the framing; its claimed detector accuracy is self-reported and not independently audited here. Spotify has not published an error rate for its persona classifier or the size and turnaround of its review team, so the badge's precision is unknown. The Chinese rule cited is in force, but platform-level practices at Tencent Music and NetEase Cloud Music are described from company statements and press reporting rather than from audited disclosures, and enforcement intensity in practice is not something these sources quantify. Figures describing a "June peak" or "up to 85%" are company-rounded statements, not measurements with a published method. This article does not cover the separate question of whether training on copyrighted recordings requires consent in Chinese courts.
What readers can do now
- If you release tracks, keep the receipts: prompts, generation logs, session files and revision history. Labelling obligations and platform reviews both turn on being able to show what the human did.
- Treat playlist inclusion as the deliverable, not output volume. With roughly 1–3% of plays going to fully machine-made tracks, an extra thousand uploads is worth less than one placement you can defend.
- Check the label rules of every platform you distribute to before you upload, and do not strip metadata. In China, removing or hiding a label is expressly prohibited, and platform penalties run from takedown to account action.