I write and record some rock and electronic music, have written and played with human musicians for decades now, and have looked at some AI collaboration tools (some through Logic Pro and some through LLM generators). I have some mixed feedback.
The AI tools are problematic by being too eager to please and not good at injecting new and contrasting ideas. A collaboration with a human is interesting specifically because the human can surprise you with something unexpected and something that you dislike at first, then convince you to keep it in until you finally do admit, yeah, that actually is way better. The AI tools (like Drummer in Logic Pro) are built to follow exactly in the direction you’re already going in. This is fine if you just want a generic backup band, but I’ve recently even stopped using Drummer (shout out to the best auto-setting, Logan), and instead record first takes to a plain click before adding my own drums mouseclick by mouseclick.
I feel like it’s a mirror for the problem with self-training. With self-training an LLM gets stuck in a little local-optimum loop where it reinforces things it shouldn’t for no clear reason, and LLM music tools tend to reinforce too strongly what you’re already doing, instead of injecting new directions and experiences.
LLMs work well for coding because code has an objective goal: convert input to output. It needs to hit the spec and pass the tests. With music there is no spec and tests, there’s just the output, and a machine telling you to do exactly what you’re doing is for most of us a direct route to a boring result.
It’s why writers need editors, and why musicians need collaborators. We need some friction in the system and for someone to tell us "No, I don’t like that, here’s what I’d do". We need someone to disagree with us.
Timestamp me though, this is 2026. It might be better in a few years/months/weeks, at the rate these things are changing.