Over the last year or so, music has probably been the area where I’ve experimented with AI the most.

At first, I wanted to see how far the technology could go. I tried different tools, voices and production approaches. Some results surprised me. AI can make a good song from a few lines of text. Sometimes a very good one.

In April 2026, I started Sizelle as an experiment: an artist project where I made and released music with AI at the center of the process. I wanted to understand what the technology could do when the result was a finished song, released publicly and listened to alongside everything else.

That meant making choices about songs, lyrics, arrangements, production and what was worth releasing. But the better the technology became, the less interested I was in making something simply because AI could make it. I became more interested in a different question: where does AI help someone create what they wanted to create, and where does it make the important creative decisions for them?

AI-generated and AI-assisted are not the same thing

Separating vocals from instruments, testing a vocal, making an AI demo later recorded by musicians, and generating a finished song from one prompt can all end up described as “AI music.” So can 10,000 songs generated and uploaded automatically.

Those aren’t equivalent. AI can be a tool, a collaborator or the source of substantial parts of the work. At the extreme, it becomes an autonomous content factory.

Electronic music makes this distinction particularly interesting to me. Drum machines, samplers, synthesizers, sequencers and software changed what musicians could make without removing the musician from the process. AI can continue that progression, but it can also go much further.

Can AI make art?

Imagine I write a melody, develop the chords, change the structure, reject most of what AI suggests, rewrite the lyrics, direct the vocal and keep changing the production until the track sounds the way I want. I have no difficulty calling that creation, even if AI was one of the tools involved.

Now imagine I type:

“Melancholic progressive house song, female vocal, emotional chorus about losing someone.”

I click Create. Two songs appear. I pick one and upload it.

What exactly did I create? The idea? Yes. The prompt? Obviously. But did I compose the melody, choose the harmony, arrange it or produce it? Mostly, I asked a machine to make those decisions.

I’m increasingly skeptical that clicking Create makes me the artist.

The resulting song could still be art. Someone could hear it and be moved by it. It could become attached to an important moment in someone’s life. Whether something is art and whether the person who prompted it is the artist are two different questions.

There is a legitimate counterargument. A film director doesn’t operate every camera. A producer may direct musicians rather than perform. Conceptual artists sometimes create work primarily through an instruction. So manual labor can’t be the test. Neither can the length of the prompt.

The better question is where the meaningful expressive decisions came from. Someone could use AI through hundreds of iterations, choices, edits and interventions and exercise enormous creative control. Someone else could write a detailed 500-word prompt, accept the first output and change nothing. The first may involve much more authorship than the second.

Copyright law can’t tell us what art is, but there is an interesting parallel. The U.S. Copyright Office said in a January 2025 report that human-authored expression can be protected even when a work includes AI-generated material. Creative arrangements and modifications can qualify, but protection doesn’t automatically extend to the AI-generated elements. The report concluded that, with the technology available at the time, prompts alone don’t establish human authorship over the generated output.

The question isn’t whether AI touched the song. It’s whether AI helped execute someone’s creative intent or supplied most of the creative intent itself.

We are heading toward unlimited music

Deezer says daily fully AI-generated uploads rose from around 10,000 at the beginning of 2025 to close to 90,000 by June 2026. On peak days, more than half of all new music delivered to the service was fully AI-generated.

Yet Deezer said those tracks accounted for around 1–3% of its total streams. Part of that gap is Deezer’s own doing, since it keeps fully AI-generated tracks out of recommendations. The company also said that, depending on the month, up to 85% of streams on those AI tracks in 2025 were fraudulent.

We clearly don’t have a shortage-of-music problem. Until recently, even a track made entirely on a laptop required someone to make hundreds of decisions. Now the marginal cost of making another song is approaching zero. When a finished song can be generated in seconds, there is little stopping someone from making ten songs, a thousand songs or a million songs.

Some completely AI-generated music will be bad. Some may be great. But if machines can produce unlimited technically competent music, technical competence itself becomes less valuable.

Production stops being scarce.

Attention is scarce.

Taste is scarce.

Trust is scarce.

Having something worth saying is still scarce.

That’s what concerns me about AI slop. Not that AI can make bad music. Humans have made plenty of that. The problem starts when content becomes almost free to manufacture at industrial scale, while the systems distributing attention and royalties weren’t designed for unlimited supply.

Should one person be able to generate a million tracks and have them compete economically with everyone else? If people genuinely want to listen to them, perhaps they should. But if the model is generate, upload, manipulate streams, collect fractions of royalties and repeat, that’s not really a question about art anymore.

It’s spam.

Platforms are drawing their own lines

Beatport rejects fully or majority AI-generated tracks while allowing AI-assisted music that remains majority human-made. That’s close to how I see it: a tool that expands what a person can make isn’t the same as a system that replaces the creative process.

Spotify is also addressing identity. Its AI Persona badge, announced for rollout from mid-September 2026, is for profiles presenting AI-generated identities rather than real people. Spotify says music from those profiles will stay out of editorial and algorithmic recommendations by default, although following an AI Persona can bring its music into personalized recommendations.

The badge is about who the artist is, not how the music was made. Its optional AI Credits cover AI use inside the creative process separately.

That distinction makes sense to me because we’re dealing with several questions at once. Was AI used? How much did it generate? Who made the important creative decisions? Was the underlying material properly licensed? And is the artist I’m looking at even a real person?

One label saying “AI music” can’t answer all of those.

Consent matters

Warner Music Group’s labels were among those that sued Suno over its training data in 2024. Warner settled in November 2025 and entered into a partnership to develop licensed AI music products, with participating artists and songwriters controlling whether their names, likenesses, voices and compositions can be used.

An artist has rights in songs and recordings, but also builds a voice, style, reputation and relationship with an audience. If AI creates a new song that copies no existing song but unmistakably sounds as if a particular artist made it, what exactly has been copied?

Copyright alone may not answer that. Morally, I think the basic principle is easier. Consent matters. Compensation matters. “The technology can do it” isn’t enough. Licensed models and real opt-in systems seem more sustainable to me than excluding the people whose work and identities create the value.

What the experiment taught me

The more music I made for Sizelle, the clearer the boundary became for me between directing a creative process and receiving an output from one.

The experiment is now almost complete. The most important thing it taught me is what I want to do next.

The next project puts people at the center. The core music and the final expressive decisions will be human-made: musicians, singers, producers and engineers making the finished record. I’ll say more about that project when there’s something to hear.

I absolutely intend to keep AI in the process. It just moves from the center to the edges: demos, tests and the small production tasks where it saves time or money.

If I’m working on a house track and wonder whether a viola or cello would work in the breakdown, I don’t necessarily need to book a musician just to answer that question. I can make a quick AI demo, hear whether the idea works, and if it does, have a real musician record it.

The same applies to voices. I can test a melody with different ranges, tones and delivery styles before deciding who should sing it, or try an arrangement that would otherwise be too expensive just to hear once.

I’m interested in Suno Studio as a new kind of DAW, with generation alongside editing and production tools. What I want from it is not “Make me a song,” but “Help me make the song I have in my head.”

AI can expand who gets to participate in making music, and I think that’s a good thing. I don’t want AI out of music. I want to see what happens when talented people get tools that let them do things they couldn’t do before.

The interesting question isn’t whether AI can make music. It clearly can. It’s whether AI helps us express something we wanted to express, or simply gives us something to publish.

And sometimes the best creative choice may still be not to use AI at all.