Constraint-Based Generative Composition
AI Music Generation is here. Composition hasnโt arrived yet.
Iโve been trying to understand something about AI music.
Not in theory. In practice.
Because I didnโt come to this out of curiosity. I came to it because I wanted a working system.
A way to keep writing songs without always going back into a recording studio. Without long sessions, logistics, favours, or the slow drain of energy that used to be part of the process. My hearing isnโt what it was. My tolerance for friction is lower than it used to be. But my interest in songwriting hasnโt gone away.
So the question became very simple:
Can AI give me a low-friction way to write and realise songs?
The answer is: yes โ but not in the way I expected.
The Assumption
I assumed that something already existed โ or was just about to.
A system where I could say:
here is a chord progression
here is a lyric
here is a melody (even roughly)
here is a structure
โฆand the AI would help me develop that into a finished track.
Not replace it. Not reinterpret it beyond recognition. Not generate something โin the style of.โ But work with it.
Suggest. Arrange. Orchestrate. Extend. Explore variations.
In other words:
behave like a very capable musical collaborator.
What Iโve come to call:
constraint-based generative composition.
What Exists Instead
What actually exists right now is something different.
Tools like Udio or Suno are impressive. In some ways, astonishing.
You can give them:
a lyric
a mood
a genre
a vocal type
โฆand they will generate something that sounds like a finished song.
Sometimes very good.
Sometimes uncannily close to something you might have written.
But there is a catch.
You cannot reliably say:
keep these exact chords
follow this melody
respect this phrasing
preserve this structure
Even if you specify it clearly.
Even if you spell it out.
Even if you know exactly what you want.
The system doesnโt work that way.
The Mismatch
And this is where the tension comes from.
Because for most songwriters, a song does not begin as a finished production.
It begins as something much simpler and much more precise:
a chord sequence
a melodic idea
a lyric fragment
a sense of form
a rhythm
Those elements are not optional.
They are the song.
Everything else โ arrangement, instrumentation, production โ can change. Often radically. But the underlying identity remains intact. You can strip a great song back to voice and guitar, or reharmonise it for a jazz ensemble, or rebuild it electronically, and it is still recognisably the same piece.
That is how composition works.
Current AI systems, by contrast, are not really operating at that level.
They are not working from composition outward.
They are working from audio inward.
Generation vs Composition
This is the distinction that clarifies everything.
What we have now is:
style-directed generation
What I expected โ and still think is coming โ is:
constraint-based composition
In the first model, you describe a musical world and the system generates something plausible within it.
In the second, you define the musical facts and the system helps you develop them.
Those are not small differences.
They are two completely different ways of thinking about music.
Why This Matters (Practically)
This isnโt just philosophical.
It affects what you can actually do.
I went into this wanting a simple workflow:
Sketch ideas and try different styles and voices
Develop a version into something usable
Rework older material where needed
Create a simple video and share it
All of that is now technically possible.
But the first step โ the actual songwriting โ is still more unpredictable than I expected.
You can generate.
You can explore.
You can discover.
But you cannot yet reliably say:
โthis is the song โ now help me realise it without changing its identity.โ
Where We Are (and Why)
This isnโt an accident.
The current generation of tools has largely been optimised for:
speed
accessibility
output
stylistic flexibility
Which makes sense.
Thatโs where the market is.
Itโs incredibly useful for:
trying ideas quickly
generating variations
creating finished artefacts with minimal effort
But it is not yet designed for:
precision at the level of composition
Whatโs Missing
Whatโs missing is a layer.
A system that understands:
chords as constraints
melody as structure
lyrics as fixed text
form as intentional design
And treats those not as suggestions, but as boundaries.
Within those boundaries, it could be incredibly powerful:
reharmonise
re-voice
re-arrange
orchestrate
generate alternate versions
develop lyric ideas
develop melody ideas
But always in relation to the original idea.
Not instead of it.
What Comes Next
This is why I think constraint-based generative composition is the next meaningful step.
Not more generation.
Not faster output.
But better alignment with how music is actually written.
The pieces are already emerging in research and early tools.
But they are not yet:
simple
unified
affordable
songwriter-first
When that changes, something important shifts.
AI stops being:
a generator of plausible songs
and becomes:
a collaborator inside the act of composition
Where That Leaves Me (For Now)
For now, Iโm using what exists.
As a way to explore.
To generate.
To test directions.
To create finished versions quickly.
And if I ever want to bring my own guitar, my own voicings, my own phrasing back into it โ I can do that later. On my terms.
But the original expectation I had โ that I could simply hand over a song and have AI help me realise it faithfully โ isnโt quite here yet.
Final Thought
Thereโs a lot of noise around AI music.
A lot of talk about scale, output, disruption.
What interests me more is something quieter.
Not how many songs can be generated.
But whether we can stay with one song โ
and go deeper into it.
Thatโs where this gets interesting.
After thought: Slop
Randomly, from time to time, AI-generated songs can be really good. However, mainly, they are formulaic slop.
Although I have NOT done deep due diligence on this particular song, I think that Always (which I featured in a recent post) is an example of an AI-generated song that works well. It rocks. It ticks most boxes (for me).
About Always being AI-generated: it is only an assumption, and I am willing to eat my hat and be proven wrong!
To illustrate the point about AI-generated formulaic slop: as a tribute to our friendship, a close friend of mine (who is actually also very proficient musician and a good songwriter) created this song, in a few seconds, on an AI-app, on his mobile, whilst in a bar, with a one or two sentence text promptโฆ
Itโs entitled: Nic-Fucking-Briscoe
Are you at a point with a song or a musical project where moving forward means making an important decision, yet one you cannot quite pinpoint?
If youโre in that position โ where the issue isnโt lack of ideas, but knowing what actually matters โ thatโs exactly the kind of work I do in my Musical Crossroads Sessions.





cool. keep playing ๐ฅ๏ธ ๐ธ