All writing

I can't believe how good AI-generated music has got

aimusic

I was scrolling this week when I came across a track that made me stop and actually check: is this really AI-generated? It was. I’ve been thinking about it ever since.

Have a listen:

That’s not a human. No session musicians, no producer, no studio time. A model generated it — melody, arrangement, production, the lot — in roughly the time it takes to boil a kettle. And it sounds good. Not “impressive for AI” good. Just good.

How we got here

A few years ago, AI music meant slightly uncanny MIDI melodies that sounded like they’d been composed by someone who had read about music but never quite heard it. The tell was always there: the timing slightly off, the chord transitions odd, the production flat.

What changed is the same thing that changed everything else: scale and diffusion.

Modern AI music tools — Suno, Udio, Google’s Lyria, Meta’s MusicGen — are built on diffusion models and transformer architectures trained on enormous audio datasets. In simple terms:

  1. Training: The model listens to millions of tracks across every genre, tempo, and style, learning the statistical patterns that make music feel like music — how tension builds and releases, how a chorus lands, how compression and reverb shape a sound.

  2. Generation: Given a text prompt (“upbeat indie pop, female vocals, sunny afternoon feel”), the model works backwards from noise, iteratively refining a waveform until it matches the description — the same way image diffusion models sculpt a picture out of static.

  3. End-to-end audio: The newest models don’t generate MIDI and then synthesise it — they generate raw audio waveforms directly, which is why the production quality has jumped so dramatically. The “feel” of a real recording — the room sound, the subtle velocity variation, the breath before a vocal phrase — is baked in at the waveform level.

Why this one hit differently

I’ve heard plenty of AI music that’s technically impressive but emotionally inert. This track isn’t. There’s phrasing in it. The dynamics feel intentional. It has the quality of something that was shaped, not just generated.

That might be a kind of illusion — the model has learned to mimic emotional phrasing so well that the distinction between mimicry and intention starts to blur. But then again, couldn’t you say the same about a lot of human music?

Where this is going

The honest answer is: very fast, and not entirely predictably. The gap between “AI-assisted” and “AI-generated” is closing. Tools that let you hum a melody and get a full production back are already in public beta. Models that can extend, remix, or respond to existing tracks are shipping.

What I find most interesting isn’t the threat to musicians — though that conversation is real and important — it’s the democratisation of a creative medium that was previously gated behind expensive gear, years of training, and access to the right people. The friction is collapsing.

I don’t know exactly what music looks like in five years. But I suspect we’ll look back at this moment — stumbling across a track and doing a double-take — as the point where something quietly shifted.