AI won't kill music. It may bury us in songs nobody cares about

AI already makes up 44% of Deezer's daily uploads

Not long ago, making a convincing song meant learning an instrument, wrestling with production software, recording a voice and spending an unhealthy number of hours changing sounds that seemed identical to everyone else. Now you can describe a sad ballad or a pop-punk breakup track and get several finished options in minutes. The lyrics may have the emotional depth of an inspirational coffee mug, but the song exists, sounds produced and can be uploaded.

That ease is already visible in the numbers. Deezer receives nearly 75,000 fully AI-generated tracks every day, roughly 44% of all new deliveries to its platform. In April 2025, the figure was just over 20,000. This is Deezer's data, not a measurement of the entire music industry, but the direction is hard to miss.

Those tracks account for only 1% to 3% of listening on Deezer. The company also classified up to 85% of streams on fully AI-generated music as fraudulent and removed them from royalty calculations. Detected tracks are labeled and excluded from algorithmic recommendations and editorial playlists; they are not automatically erased from the catalogue.

The first major disruption is likely to hit music that sits behind something else: videos, ads, podcasts and smaller games. Many of those jobs do not need the song of the year. They need something suitable, unobtrusive and inexpensive. A subscription that produces 20 options in an afternoon competes directly with people who used to be paid for that work.

Taylor Swift will not be the first person hurt

Taylor Swift, Bad Bunny and Rosalía are not valuable only because they can sing or produce. Their work carries a history, an aesthetic and a community that follows what they do. An app can manufacture a competent chorus. Manufacturing years of connection with millions of people is much harder.

The earlier pressure may fall on people working around music without becoming its public face: library composers, session players, demo producers, jingle writers and commercial vocalists. New artists face a different version of the same problem. Releasing a song gets easier while earning attention among millions of cheap tracks gets harder.

In a blind test commissioned by Deezer and run by Ipsos, 9,000 participants heard two AI-generated songs and one human-made track. Ninety-seven percent failed to classify all three correctly, yet 80% wanted fully synthetic music to be labeled. Our ears may not always know the difference, but people still care about who—or what—made what they hear.

That is why concerts, interviews, studio footage and direct relationships with fans will matter even more. They cannot make a weak song good, but they offer something a content farm struggles to copy: a person worth following.

The record labels went from lawsuits to negotiations

In June 2024, labels affiliated with Universal, Sony and Warner filed separate lawsuits against Suno and Udio. They alleged that the companies copied protected recordings to train their models without permission. Those were claims awaiting judgment, not court findings.

Then came the deals. Universal settled with Udio in October 2025 and announced a future service trained on authorized music. The following month, Warner struck its own Udio agreement and settled with Suno. Warner artists and songwriters will be able to opt in before their names, images, voices or work are used in new songs, according to the terms announced by the company.

The labels want control, but they also see a business. A famous voice can be licensed for personalized songs, remixes or collaborations that would never enter a traditional studio. The real fight will be in the contract: how much the artist earns, which uses they can refuse and who stops a creation they would never have approved. Consent is already a serious problem in apps that manufacture fake intimate images; a famous voice raises a similar dispute.

Musicians will have more tools—and far more competition

Music AI is not limited to generating finished tracks. It can separate instruments, clean recordings, test harmonies, draft arrangements or turn a phone recording into a demo that communicates an idea to the rest of a band. In the hands of someone with taste and experience, it can remove hours of repetitive work.

What it cannot settle is whether an idea deserves to become a song. Anyone can generate 100 choruses. Choosing one, giving it a point of view and knowing when to delete the other 99 still requires judgment. The technical barrier falls just as the competition for attention rises.

We will get personalized songs, virtual artists, licensed voices and music that changes with what happens inside a game. Some of it will be useful, some will be fun for five minutes and an enormous amount will pass unnoticed.

There will be no shortage of music. The hard part is making any of it matter.