Generative music makes more than you put in, out of rules a person wrote, with no AI and no training data anywhere near it. Here is the real machinery: the handful of old, well understood techniques that let a few rules play forever.
Enough of the how that you can spot a generative tool, and see why it is a different thing from AI.
Constrained choice. That is the whole trick.
Pure randomness is noise, every note equally likely, and it sounds like it. A fixed sequence is a loop, the same thing every time, and it gets old. Generative music lives in between. At each moment there is a set of allowed options and a weighting on them, and the machine makes a roll inside those limits. The rules keep it musical, the roll keeps it moving, and the same rules give you something a little different every run.
The most common engine, and simpler than it sounds. A Markov chain is a table of "given where you are now, here is what tends to come next, and how likely each option is". Land on a certain chord and the table says the next one is probably this, sometimes that, rarely the other. You roll against those odds and move. Do it over and over and you get a line that makes sense step to step without ever being a fixed melody. Feed the table musical odds and it wanders like music, not like a dice cup.
Odds alone are not enough. The thing that stops generative music sounding like a clever accident is hard musical constraint layered on top. Stay in a key. Put chord tones on the strong beats and passing notes between them. When chords change, move each voice the shortest distance, which is the trick called voice leading and the difference between a keyboard demo and something that sounds composed. These are rules a theorist could write on a napkin, and they carry most of the musicality.
Music is not just a stream of good next notes, it has shape. So generative systems borrow ideas from grammar: a phrase, an answer to it, a return. Above all, repetition. A cell that comes back is what turns motion into a hook, and a good generative patch repeats on purpose rather than forgetting what it just played.
All of it is nailed to a tempo and a grid. The cleverest note choices in the world do nothing if they do not land in time, so the timing is usually the least random part of the whole machine.
Read back over that list. Every piece is a rule a person wrote and could explain to you. There is no model, no neural network, no dataset of other people's songs, nothing learned. The variation comes from a roll of the dice inside human-set limits, not from a machine predicting what a million tracks would probably do next.
That is the clean line between generative and AI, and it matters more than it used to. If you want it laid out properly, we wrote a whole guide on generative music versus AI music.
This is exactly how ours work, and none of them touch AI. ARGISH self-plays a chord progression with a Markov chain on a chord wheel, then voices the pads with voice leading so they never jump. REHEAT writes an acid bassline as a repeating cell with the accent and slide grammar of a 303, locked to your kick. SILT runs four tape loops of coprime lengths so they drift out of phase and the music writes itself from the overlap. Rules, a clock, and a bit of controlled randomness, doing what this guide describes.
Easiest to just hear it. All three play free in the browser.
We write these when there is something worth writing down. One email when a new one lands or a new Tunary instrument ships. No newsletter, no schedule.
Tunary