Journal · AI & production

AI and music: what machines can do, and what only a composer should

Every conversation about music now arrives at the same question, usually within minutes: will AI replace composers? I have answered it on a festival stage in Monaco, in contract negotiations with broadcasters, and at two in the morning in my own studio with the tools open in front of me. So this is the long answer - from someone who is neither afraid of the machine nor in love with it, and who uses it every working day without ever letting it write a note that matters.

Where this opinion comes from

In June 2026, AI Filmfest Monaco put me on a panel at One Monte-Carlo to talk about how to run a creative studio in the age of AI. I took the stage as someone who works on both sides of the line: I build AI-assisted workflows for brand work across Europe, the Middle East, and Asia - and I score for German and Italian television, markets whose contracts now demand written guarantees about how the music was made. That double life keeps the opinion honest. I cannot afford to be a romantic about the work, and I cannot afford to be careless about the machine.

What the machine is genuinely good at

Here is what almost every article about AI music gets wrong: the interesting use is not typing a prompt and receiving a song. The interesting use is running an assistant as a layer connected to your studio - doing the work a studio runner used to do, plus a few things nobody could do before.

In my studio that looks like four things. First, the grunt work: preparing and cleaning sessions, organizing and color-coding tracks, setting the room up before a take - the hour of friction that used to sit in front of every idea. Second, a personal sound library: capturing real sounds out in the world and bringing them straight into the session, a palette no sample pack can sell you. Third - and this is where it gets genuinely new - building instruments: turning a captured or imagined sound into a playable soft synth, a VST I perform live on a MIDI keyboard. Not choosing presets; building the instrument. And fourth, reverse-engineering timbre: having the assistant analyze a sound’s frequency response and harmonics, then rebuilding that exact character inside a synth - recreating in minutes a tone that once took hours of sine waves and guesswork.

Notice what all four have in common: the machine handles scale and friction, and every musical decision - every note, every sound that ends up in the score - still passes through human hands. AI handles the scale. I handle the meaning.

What the button gets wrong

Then there is the other kind of AI music: the generator. Type a brief, receive a track. And I will say something here that composers are not supposed to say - for genuinely functional music, it can work. If success means “90 seconds, upbeat, inoffensive, background,” the machine can hit that brief, because the brief is fully specifiable. That music was wallpaper before AI, and it is wallpaper now; only the wage changed.

Film scoring is a different job. You are not serving a brief; you are serving a director’s interior vision - often something they cannot fully put into words. “The middle should ache more.” “It needs more sunrise.” The difference lives precisely in the gap between a specification and a feeling, in translating narrative subtext into harmony (the mechanics of which I’ve written about in scoring emotion, not scenes). A generator cannot stand in that gap, because there is nothing to prompt with - the whole point is that no one has found the words yet.

And there is a subtler danger for the people who make things. Unguided, the machine doesn’t just assist your creativity - it tilts it. It offers, you accept, it decides, you follow. Fifty small conveniences later, the work sounds like the model’s taste instead of yours. The discipline is in refusing that trade, every day, one decision at a time.

How music generators actually work - and why provenance matters

To understand both the promise and the problem, you need one mental model: a music generator does not compose, it continues. These systems are trained on enormous catalogs of existing recordings, learning the statistical texture of music - what tends to follow what, how a chorus lifts, what “cinematic” sounds like on average. When you prompt one, it produces the most plausible continuation of your request through everything it has absorbed.

Two consequences fall straight out of that design. First, the average problem: because the model learned what usually happens, its instinct is the middle of the road - competent, familiar, and interchangeable. That is fine for wallpaper and fatal for identity; a brand asset that sounds like the statistical average of its genre is, by definition, not distinctive. Second, the provenance problem: the model’s entire ability is distilled from other people’s work. Whether that distillation was licensed is not a philosophical question - it is the question, and it is why the training-data lawsuits happened, why licensing deals are now being signed, and why professional buyers have started asking where the music in their campaigns actually comes from.

This is also why the “it’s just a tool, like a synthesizer” argument only half works. A synthesizer generates sound from oscillators; it owes nothing to anyone. A generative model generates music from music; it owes everything to its training set. The tool analogy holds for how I use AI in my own studio - assistants, analysis, instrument-building - and breaks precisely at the point where the machine starts producing finished musical material of uncertain parentage.

Will AI replace composers?

No - but it will replace music that never needed a composer. Every wave of technology in this field has triggered the same funeral: the drum machine was going to kill drummers, sampling was going to kill orchestras, the DAW was going to kill studios. What actually happened, every time, is that the tool absorbed the functional work and raised the value of the irreplaceable part - taste, accountability, and a person in the room who can be trusted with a feeling. The hybrid orchestral language that dominates modern cinema is itself the child of one of those funerals: the orchestra didn’t die when the synthesizer arrived, it merged with it.

What AI cannot deliver is the thing clients are actually buying: someone answerable. A composer hears the note that is wrong and knows why. A composer sits with a director at 2 a.m. and finds the version that aches correctly. A composer signs their name to the work and carries it. There is no model output you can hold accountable, and in professional work, accountability is not a soft skill - it is the product.

The legal reality brands cannot ignore

Even if you set the artistry argument aside entirely, the commercial one remains - and it is sharp. As of this writing, the legal ground under generated music is still moving: the major labels sued the big music generators over training data, some of those suits have settled into licensing deals, and the core fair-use question is still before the courts. But one fact matters more than all the litigation for anyone building a brand: purely AI-generated work is, under current US guidance, not copyrightable. Nobody owns it - which means you cannot own it. You cannot stop a competitor from using something functionally identical, because there is nothing to enforce.

For a sonic identity, that is disqualifying. The entire value of a sonic brand asset is exclusive, compounding recognition - a sound that belongs to you and no one else, the argument at the heart of custom music vs. library music. An asset nobody can own is not an asset. And broadcasters have already priced this in: the German and Italian television markets I write for now ask composers to guarantee the provenance of their music in writing. The industry is not waiting for the courts; it is contracting around the risk today.

The three questions

On that stage in Monaco I proposed a standard I would happily see the whole industry adopt - three questions before any AI tool touches professional work. Where was this model trained? If you don’t know what it learned from, you don’t know whose work you are borrowing. Do I legally own what it produces? If the answer is unclear, so is your client’s asset. And the hardest one: could I have made something equally good without it? If you cannot answer all three, you are not using a tool. You are outsourcing your judgment.

Using AI in music ethically: a working checklist

For the musicians and studios asking how to actually operate in this landscape, here is the standard I hold my own studio to - five rules that have survived contact with real client work.

Keep authorship at the decision layer. The machine can prepare, analyze, and build tools; the notes, the arrangement, and the final sound are chosen by a person who can explain why. If you cannot narrate the musical decisions in your own work, you didn’t make it. Know your chain. Every tool in the signal path should have an answerable origin - what it was trained on, what its license says, what its output status is. Never launder someone else’s work. Style-transferring a living artist, cloning a voice without consent, or generating “in the style of” a catalog the model ate without permission is theft with extra steps. Disclose where it matters. Clients, broadcasters, and collaborators deserve to know how the music was made - especially now that contracts ask. And ask the three questions - every time, without exception, including on deadline. Especially on deadline.

The palette, not the point

I was classically trained on a violin. I learned production from the waveform up. And I now run an assistant wired into my DAW that does in minutes what once cost me nights. None of these facts contradict each other - they are the same career, because the standard never changed: new tools enter the palette only when they serve the story. The synthesizer earned its place that way. AI is earning a narrower one, further from the meaning and closer to the friction, which is exactly where it belongs.

Wherever the music comes from - if there is no meaning behind it, it’s just noise.

Frequently asked questions