Forty-four percent of new music uploaded to Deezer is fully AI-generated, according to figures released in July 2026 by a music-industry labelling initiative. Apple Music puts its own figure at more than one-third. These are not assistive tools helping human producers polish a vocal take or generate a drum pattern. These are tracks where the composition, performance, and production originate from a prompt and a model, with no human hand on the instrument. The platforms are counting these uploads, not celebrating them.
The test no one asked for
Can anyone hear the difference? A blind listening test followed by disclosure offers a clean way to find out. Listeners rate tracks cold, on nothing but what reaches their ears. Then they learn which songs were made by humans and which by algorithms, and rate again. The gap between the two scores measures something the music itself cannot contain: our need for the story behind the sound.
Early trials of this format, run by researchers and journalists outside South Africa, have produced a consistent pattern. In the blind phase, listeners often rate AI-generated pop, ambient, and electronic tracks as competent, sometimes even pleasing. The melodies hold. The production is crisp. After disclosure, the same tracks drop. The notes did not change, but the listener’s frame did. Knowing a song is AI-generated triggers a reassessment of what the song is for. We hear intention, struggle, choice, a person standing behind the decision to move from verse to chorus a beat earlier than expected. Strip that away and the same waveform becomes functional noise, a utility, something closer to hold music than to art.
This bias is not irrational. It is the correct response to a category error. Most AI uploads are successful attempts at volume, not failed attempts at beauty.
Two kinds of uploaders
The flood of AI music splits into two streams that share nothing but their production method. One is genuine experimentation: producers using generative models as instruments, curating output, making aesthetic choices, treating the machine as a collaborator with known limitations. The other is royalty farming, the mass production of generic tracks designed to capture micro-payments from background listening, algorithmic playlists, and sleep-study-meditation channels.
The economics are brutal and simple. A human artist might spend months on an album, pay for studio time, hire session musicians, and release ten tracks that earn streaming royalties split among labels, distributors, and platforms. An AI operation can generate a thousand tracks in a weekend, upload them through automated distribution services, and collect fractional cents across enough plays to turn a profit at scale. The content does not need to be loved. It needs to be played, briefly, by someone not paying attention.
The Deezer and Apple Music figures are not signs of creative flourishing. They are signs of an arbitrage opportunity being exploited faster than platforms can respond. The 44% figure is 44% of what got uploaded, not 44% of what people searched for.
What labels can and cannot fix
Music organisations have proposed a labelling system to distinguish AI-generated recordings from AI-assisted work. This distinction helps listeners understand what they are supporting, and it helps rights holders trace who owes whom. A track where a producer used AI to suggest chord variations, then selected, arranged, and performed the result, retains human creative decision at its centre. A track generated from a text prompt and uploaded unchanged does not.
Labels, however, solve only the transparency problem. They do not solve the economics problem, the discovery problem, or the quality problem. A listener who sees an “AI-generated” tag on a track they already enjoyed in a blind test faces a small crisis of category. Do they keep listening? Do they feel deceived? Do they stop trusting their own ears? The label gives them information, but it does not tell them what to do with it.
Platforms face the harder structural question. They built business models on near-infinite content catalogues, where scale was a feature. Now scale has turned against them. Storing, moderating, and surfacing millions of AI-generated tracks costs real money in server infrastructure and curation labour. The micro-royalties paid to AI uploaders come from the same pool that pays human artists. Every rand directed toward a generated ambient loop is a rand not directed toward a composer who rented studio time in Braamfontein or Cape Town.
The real question
The central issue is not whether AI can make music; it clearly can. The issue is whether the current volume of AI-generated content corresponds to any listener demand, or whether it is purely supply-driven, an artifact of cheap generation meeting a payment system designed for scarce human labour.
The evidence points toward supply. Listeners do not search for “AI-generated lofi beats.” They search for “lofi beats,” and algorithms fill the playlist with whatever meets the acoustic profile at the lowest acquisition cost. The listener gets what they asked for functionally, but not what they might have wanted relationally: a connection to a person who made a choice, took a risk, had a bad day, recovered, finished the track anyway.
This is the gap the blind test exposes. We are not bad at detecting AI music. We are bad at wanting it once we know what it is. The technology has solved generation. It has not solved desire. The 89,999 tracks uploaded alongside one genuine experiment do not fail because they sound wrong. They fail because they mean nothing, because no one made them for any reason except to exist and to earn. We can hear that, or rather, we can feel its absence, once we know to listen for it.
The platforms will eventually throttle, label, or filter this content. Until then, the percentage will keep climbing, and the test will keep producing the same result: a song is a relationship between a listener and the knowledge that someone, somewhere, meant it.
