AI tools can now generate vocals, instruments and complete arrangements from a short written prompt. But listeners aren’t interested. Credit: Campus / Shutterstock
AI music is taking the music industry and platforms by storm. Its melodies are flawless, can give a moody vocal performance that resembles Billie Eilish, and can create a song in pretty much any style the prompter likes. With the vast choice of platforms people now have to upload their own music daily, anyone can create a fake AI persona, make a few songs with services like Suno or Udio, and put them up on Spotify and other streaming services.
The robots have found the upload button
That freedom to upload, with no consistent or visible categorisation on whether a song is written or made with AI, is what’s sparking massive debate across the music industry. While artists are furious at the lack of labelling or control over how AI uses human artists’ songs to train its models, general listeners are also becoming less and less capable of identifying whether a track they’re listening to was created by an AI or a human. As general opinion goes, one could say it’s down to the listener whether they want to listen to a machine speak about love or grief with a perfect note-for-note vocal performance, or whether they’d prefer to hear the natural imperfections and emotion from a singer while their quivering lip sings of the loss of a loved one.
However, that being said, before anyone starts imagining robot stadium tours, the headline needs a little explaining. Deezer, a French music-streaming service similar to Spotify or Apple Music, said on July 21 that it was receiving almost 90,000 fully AI-generated tracks every day. On peak days in June, they made up more than half of all new music delivered to the service. Of course, that is more than half of the Deezer platform’s new uploads, not half of all the music uploaded in the world to every platform. It is an important distinction, although 90,000 machine-made songs arriving daily on just one service is hardly a modest pile. And nor is Deezer alone. The Recording Industry Association of America said Apple Music had separately reported that more than one-third of tracks uploaded to its platform were entirely AI-generated.
The rush of new material doesn’t seem to be producing a matching rush of fans though. Fully AI-made music accounts for only a measly 1 to 3 per cent of listening on Deezer, and the company found that up to 85 per cent of the streams generated by those tracks during 2025 were fraudulent and therefore excluded from royalty payments. Deezer says it keeps detected AI tracks out of its recommendations and editorial playlists, and it now intends to remove those connected to streaming fraud, along with AI tracks that go six months without a single play.
The question that then arises is, if hardly anyone is listening to the music being made by AI and uploaded, who is making it? And who is it for? Because so far, it looks less like a public clamouring for robot pop stars and more like a content factory leaving the tap running.
Most listeners can’t hear the trick, which makes it more uncomfortable
AI artists may not have won over many fans yet, but they’re becoming surprisingly good at sounding as though somebody is actually there. In a Deezer-commissioned Ipsos survey, 9,000 adults across eight countries, including the UK, listened to three recordings: two fully generated by AI and one made by a human. Some 97 per cent failed to identify all three correctly. One caveat is that this doesn’t mean every participant mistook every AI song for a human recording. It means only 3 per cent classified the complete set correctly. Still, it’s hardly a ringing endorsement for our ability to differentiate the two.
More than half felt uncomfortable about the result, while 80 per cent supported the idea that there should be clear labels for fully AI-generated music on streaming platforms and elsewhere. Some 52 per cent also said they thought synthetic songs shouldn’t compete in the same main charts as human artists.
The figures expose a real gap between convincing sound and meaningful art. AI’s song-making capabilities are so advanced now that they can reproduce the trembling vocal of a singer, a sad piano and a carefully timed swell at the end of a chorus. But emulating is easy for a machine. Living a life with real emotions, having real experiences, and then expressing them in a song for people to connect with is rather harder.
Musicians are not declaring war on every clever tool
The argument is sometimes presented as musicians against technology, which is far too basic an argument. Artists already use software for audio repair, stem separation, mastering, vocal processing and countless other jobs that once took much longer. New software over the years, including the recent AI developments, has allowed millions of artists all over the world to share their music from their home. Once upon a time this would only have been possible for a lucky few who had access to money, connections in the industry, or by spending their entire monthly salaries on a chance at some time in a professional studio.
However, the line becomes sharper when a short prompt produces the lyrics, voice, instruments, arrangement and finished recording. Composer and flautist Daniel Finch, writing for the Musicians’ Union, described generative AI “creators” as “not true creatives, they are curators”.
UK musicians also appear remarkably united on the basic protections. A UK Music survey found that 92 per cent wanted AI-generated songs labelled. Ninety-one per cent said consent should be obtained before recordings and compositions are used to train AI models, while 93 per cent believed AI companies should pay for that material.
Honest labels could stop the guessing game and provide choice
Music organisations including the International Federation of the Phonographic Industry, the Grammys and the Recording Industry Association of America have now proposed two voluntary labels. “AI-Generated” would cover recordings where AI produced all or most of the creative elements, including a lead vocal or key instrumental performance. “AI-Assisted” would identify substantially human work that used generative AI for certain elements.
The difference is debatably quite big. Tidying a noisy vocal and inventing the singer are not quite the same afternoon’s work. Especially when the creator is making up an entire social media persona for a fake AI artist alongside it.
The scheme is also voluntary and does not yet cover AI-generated lyrics, compositions, videos or cover art, which means a huge gap would still be present in the value people perceive these aspects of songwriting and production have. This then leaves the same important question open once again: should labelling and crediting be obligatory?
European rules will begin applying on August 2. Under Article 50 of the EU AI Act, providers of generative AI systems must make synthetic audio machine-readable and detectable. But that technical marking does not necessarily mean a clear badge will appear beside every song.
For now, labels won’t settle whether fully synthetic music belongs on streaming platforms. Only people can decide what they want to listen to, and it should remain that way. But at least they would let people know what they are hearing before an algorithm sneakily makes that choice for them.