In July 2026, Deezer reported that AI-generated tracks had exceeded half of all daily music deliveries at peak levels in June: approximately 90,000 tracks per day.
The figure sounds extraordinary. But before concluding that half of modern music is made by artificial intelligence, we need to ask a more important question: what exactly does a platform count as AI music?
Fully generated recordings? Synthetic vocals? AI arrangements? Individual generated instruments? AI mastering? Restoration? Reconstruction of a human performance? Or any track that passed through a machine-learning tool?
The answer changes the meaning of the statistic.
What Deezer Actually Measures
Deezer's latest figures do not cover every recording made with any form of AI assistance. They refer specifically to fully AI-generated or fully synthetic tracks identified by its detection system.
In June 2026, Deezer reported approximately 90,000 fully AI-generated tracks arriving each day, exceeding 50% of daily deliveries at peak levels. Yet these recordings represented only around 1–3% of total streams, and up to 85% of their streams were identified as fraudulent. More than 13.4 million AI-generated tracks were detected and tagged in 2025. (Deezer, July 2026)
These are important figures. But they describe the mass supply of fully synthetic content—not the total use of AI throughout music production.
Deezer analyses audio for signatures associated with generative models. Its system initially targeted major generators including Suno and Udio, while later work sought greater generalisation to unseen systems. (Deezer AI labelling) Deezer researchers reported accuracy as high as 99.8% on test data, while warning that a strong test score does not solve every real-world detection problem. New models appear, audio is altered, and hybrid workflows complicate classification. (Deezer Research)
The responsible interpretation is therefore:
More than half of new daily deliveries at peak levels were identified by Deezer as fully synthetic recordings.
It does not mean that AI was used in only half of all uploaded music.
“AI Music” Is Not One Category
Public discussion uses AI music as a single label for very different processes.
| Category | What happens | Usually detectable from final audio? |
|---|---|---|
| Fully AI-generated | Nearly the entire recording, including vocals and instruments, is generated | Sometimes, if the detector recognises the model |
| Primarily AI-generated | AI creates most of the recording; a human selects, edits or adds material | Possibly, but the boundary is less reliable |
| AI-generated vocal | A synthetic or cloned voice is used | Sometimes |
| AI-generated instrumentation | AI creates individual parts or the accompaniment | Sometimes |
| AI-generated composition | AI creates melody, harmony, structure or musical ideas | Not necessarily; humans may perform the composition |
| Hybrid human–AI production | Human and generated performances coexist | Partly or not at all |
| AI-assisted production | AI supports arrangement, editing, sound selection or production decisions | Usually not |
| AI-reconstructed performance | AI reconstructs or extends material based on human performance | Not always |
| AI processing | Stem separation, restoration, denoising or enhancement | Usually not |
| AI mixing/mastering | AI supports mixing, mastering or technical optimisation | Usually not |
| Human re-performance of an AI idea | Humans record material first proposed by AI | Almost impossible to determine from the final audio |
A binary question—“Is this AI music?”—cannot adequately describe this spectrum.
The Invisible Area: AI-Assisted Music
Consider a song whose lyrics, melody and vocal are human; whose original instrumental performances were played by musicians; and in which AI was used to reconstruct or transform part of the arrangement while selection, structure, editing and artistic direction remained human.
That work is neither fully human in a technical sense nor fully AI-generated. It is human-led, AI-assisted production.
It is also unlikely to appear in statistics designed to detect fully synthetic audio.
The same applies to the growing amount of music using AI for vocal cleanup, noise removal, stem separation, restoration, pitch and timing correction, sound search, additional layers, mastering, format adaptation or mix analysis.
There is therefore a major difference between:
- tracks a platform can identify as fully AI-generated; and
- tracks whose production involved AI at any stage.
The first number is beginning to be measured. The second remains largely unknown.
Detection and Disclosure Measure Different Things
The industry currently obtains AI information through two main mechanisms.
Detection analyses the finished audio for evidence of generation. It does not depend on an uploader's declaration, but it can miss unknown models, heavily edited output, hybrid recordings, AI-originated compositions re-performed by humans, and technical AI processing that leaves no recognisable generative signature.
Disclosure relies on creators, labels or distributors submitting metadata about where AI was used. Spotify supports a more granular, role-based approach developed through the music metadata ecosystem: AI credits can identify lyrics, vocals, instrumental performances and production separately. Human and AI contributions in the same role can also be credited separately. The credits apply to a contribution, not automatically to the entire track, and disclosure is currently optional. (Spotify: AI credits)
Spotify explicitly recognises that AI use is a spectrum rather than a binary choice, and says disclosure is not intended to punish responsible AI use or down-rank a track. (Spotify on AI disclosure)
But voluntary disclosure is incomplete too. Distributor forms differ; older catalogues lack the metadata; definitions are interpreted inconsistently; and many technical AI tools are treated as ordinary software rather than declared creative contributions.
Detection tells us what a system could identify. Disclosure tells us what a rightsholder chose and was able to report. Neither reveals the whole market by itself.
Why Industry Figures Cannot Be Compared Directly
The percentage of daily uploads is not the percentage of an entire catalogue, and it is not the percentage of listening. A generator can supply thousands of tracks that attract almost no genuine audience.
The percentage of streams measures consumption, but Deezer's figure is complicated by the unusually high level of fraud detected in fully generated catalogues.
The percentage of revenue is different again. A CISAC/PMP Strategy study projected that GenAI music could account for roughly 20% of traditional streaming-platform revenues and around 60% of music-library revenues by 2028. This is an economic projection of future GenAI output penetration—not a count of existing tracks. (CISAC/PMP Strategy AI Study)
Finally, surveys of musicians measure the use of tools: ideas, lyrics, sounds, mixing, mastering or administrative assistance. That is production-practice data, not a count of fully generated recordings.
“Half of uploads are AI-generated” and “half of musicians use AI” could both be correct while describing entirely different realities.
What the Major Sources Can—and Cannot—Tell Us
| Source | What it can show | Main limitation |
|---|---|---|
| Deezer | Recordings detected as fully AI-generated; upload and stream shares | Does not measure the full range of AI-assisted production |
| Spotify | Role-specific AI credits delivered by partners | Optional disclosure; no complete public catalogue statistic |
| Distributors | User declarations at upload | Forms and definitions differ |
| DDEX-based metadata | Potentially granular descriptions of AI contributions | Adoption remains uneven |
| CISAC and industry studies | Economic projections and creator impact | Modelling rather than direct track counts |
| Third-party detectors | Probability of synthetic origin | No unified official methodology |
The absence of a published figure from Spotify, Apple Music, YouTube Music or Amazon Music does not demonstrate an absence of AI music. It means that no public statistic with an interpretable methodology has been supplied.
Where the Definition Still Breaks Down
Is automatic noise removal AI music? Is an algorithmic mastering suggestion AI-assisted production? If a model creates an instrumental foundation that musicians completely re-record, is the finished recording AI-generated—or is only the composition history connected to AI? If AI reconstructs a performance from human source material, should the disclosure concern performance, production or processing?
A musical release exists on several levels at once: composition, lyrics, performance, sound recording, arrangement, production, post-production, visual presentation and artist identity. AI may participate in one and not the others.
Spotify, for example, stresses that an AI Persona badge concerns an artist's public identity, not how their music was made. An AI-generated artist image and an AI-generated sound recording are different classifications. (Spotify: AI Personas)
What More Honest Statistics Would Require
Instead of one universal “AI music” category, useful statistics should distinguish:
- fully AI-generated recordings;
- primarily AI-generated recordings;
- hybrid recordings with generated elements;
- AI-generated composition or lyrics;
- synthetic or cloned vocals;
- AI-assisted production;
- AI-reconstructed or processed material;
- AI mixing and mastering; and
- unknown or undisclosed use.
Each level should then be separated by uploads, total catalogue, genuine streams, unique listeners, revenue, fraudulent streams, and editorial or algorithmic distribution.
Only then could the industry discuss AI's share with real precision.
What We Can Responsibly Say Today
We can say that mass delivery of fully synthetic music is growing extremely quickly; that it exceeded half of Deezer's daily deliveries at peak levels; that upload volume has not translated into equivalent listener interest; that a large share of streams in this catalogue has been linked to manipulation; and that the industry is moving from binary labels toward role-specific credits.
We cannot yet establish what percentage of all released music uses AI somewhere in its workflow, how much human-led music uses AI mixing or processing, how many hybrid works escape detection, whether platforms classify the same recording consistently, or how complete voluntary disclosure is.
Conclusion
Current AI-music statistics are not necessarily wrong. But almost every figure is true only inside its own methodology.
The most visible part of the market—fully synthetic recordings—is beginning to be counted. Much less visible is the music in which human authorship, performance and artistic direction coexist with AI arrangement, reconstruction, processing or production tools.
That area may be much larger than official labels suggest.
The central question is therefore no longer simply:
How much AI music exists?
The more accurate question is:
Which contribution of AI are we measuring—and which forms remain invisible?