How Musicians Actually Use AI

The most consistent finding across practitioner interviews isn’t a technique — it’s a silence. “Everyone is using these tools, artists at all levels,” music-tech executive and Machine Music author Drew Thurlow told MusicRadar in April. “They don’t want to talk about it.”

Jimmy Iovine said the same thing from the other side of the industry in an August New York Times interview alongside Dr. Dre: there are “a lot of closet AI producers out there.” Dre agreed — “They’re using it, they just don’t want to admit it” — and made his own position unusually public: “I think the only people that see it as a threat are the people who have trouble creating.” He compared the resistance to earlier panics: “You sound like the person that would have been against the drum machine when it came out. Or synthesizers, right? It’s a new tool for creativity.”

The stigma is understandable — public discourse has largely collapsed into “AI slop” versus open letters, and an artist who admits to using these tools risks being filed under the former. But it means the most interesting conversation in AI music is happening quietly, in production interviews and studio podcasts rather than announcements. That’s the conversation the rest of this feature is about.

Fred Falke — one half of the French Touch duo Braxe & Falke, with production credits going back to the genre’s late-90s foundations — initially dismissed generative AI entirely. What changed his mind was watching how a friend used it: not to make songs, but to make material. Falke now approaches Suno the way earlier producers approached crate-digging — prompting for a specific drum loop or melodic phrase, then sampling and processing the output like any other source. He also uses it for a mundane but previously difficult job: “it can really remove all the reverbs, all the delay, and you have a dry vocal.”

The producers behind Charli XCX’s “No One Lasts Forever” pushed the same idea somewhere stranger. AG Cook and Finn Keane spent a studio session deliberately making bad sounds — “We were literally running around the studio with as many things plugged in as possible and just doing the worst sounds” — then fed the fragments into Concatenator, a machine-learning sampling plugin from Datamind Audio that reconstructs incoming audio in real time by stitching together matching fragments from the pile. The drum pattern on the finished track came out of that process. “It was a nice way of being able to sample tons of stuff in this really ambiguous, uncanny way,” Cook said.

What both workflows have in common: the AI never touches the song. It generates clay. The musicianship — choosing, editing, arranging, discarding — stays exactly where it always was.

The second recurring workflow is using AI as a second opinion on music you already wrote. Alan Braxe feeds his own instrumental tracks into generative tools and asks for rearrangement variations — not to use the output, but to hear what else his material could be. He estimates AI touches “15% or 20%” of his process; the rest remains hardware synthesizers and the same practice he’s refined for decades.

Dr. Dre’s description of his own use is nearly identical in shape: “as a tool to see what it would do with what I just did.” Not generation from nothing — reflection on something. The artist supplies the idea; the model supplies alternatives the artist is free to reject, and mostly does.

This is probably the least visible and most widespread professional use case, because it leaves no fingerprint in the finished record. Nothing the AI produced survives to the master — only decisions it helped the artist make.

The artists who have gone deepest into AI are, tellingly, not the ones licensing a startup’s model — they’re the ones building their own. Imogen Heap, rebuffing offers from AI companies, worked with an audio engineer to train ai.Mogen on her own voice, and described the moment it clicked in a remix experiment: “It was the weirdest thing, but it sounded amazing. It was my voice trying to sing the kick and snare, the bassline, the keys.” Her stated ambition is specific: “I’m not excited about AI making stuff that sounds like something a human could do — I want to hear things that humans have never imagined.” She is now building an app to let other musicians train private voice models — control, in her framing, being the entire point.

Holly Herndon got there years earlier. Her AI ensemble member Spawn was trained entirely on her own voice and data, and its successor Holly+ lets anyone perform through her voice — with her consent, on her terms. Herndon’s decade of practice is the standing proof that the technology and the artistry aren’t opposed; the question is who owns the model.

That question — not “AI: yes or no” — turns out to be the real dividing line among artists. Heap and Herndon are frequently described as AI optimists, but neither is endorsing what the anti-AI open letters attack: models trained on other people’s work without consent. They’ve simply arrived at the answer by building the consensual version themselves.

Beneath all of this sits a tier of AI so thoroughly absorbed into professional workflows that practitioners barely count it as AI anymore: stem separation, mastering assistance, mix cleanup, noise removal. These tools now ship inside the major DAWs and standard plugin suites, and their adoption long ago stopped being controversial — nobody writes an open letter about isolating a vocal from a stereo bounce. This is where AI’s professional penetration is deepest, precisely because it assists a decision-maker rather than replacing one.

At the opposite edge, some artists are pushing assistance into performance itself. Dutch techno producer Reinier Zonneveld trained a model on his own back catalog and performs live B2B sets with it — the AI listens to what he plays and responds through drum machines and synthesizers on stage. It’s a radical setup built on the same conservative principle as Heap’s and Herndon’s: the model is trained on his own work, and he’s standing next to it making the calls.

Put the workflows in a line and a structure appears — call it the assistance spectrum:

Invisible assist (stem separation, mastering, cleanup) → raw material (Falke’s loops, Cook’s Concatenator drums) → the mirror (Braxe’s rearrangements, Dre’s “what would it do with what I just did”) → own-voice models (Heap, Herndon, Zonneveld) → virtual artists (fully AI-fronted acts, like Timbaland’s Suno-built Tata Taktumi).

Almost every working musician quoted in this feature operates in the first four tiers, where a human supplies the material and makes the decisions. Nearly all of the public controversy — the open letters, the streaming-platform bans, the chart eligibility rules — targets the fifth, plus the unlicensed training data underneath the tools themselves. Those are real fights. But conflating the whole spectrum with its far end is how the discourse ended up describing a technology that most professionals use like a better sampler as an existential replacement for musicians.

The policy world is, in effect, slowly rediscovering this spectrum: every disclosure regime now rolling out — tags, badges, eligibility rules — is an attempt to ask where on the line a recording sits. The musicians were already keeping score. They just weren’t saying it out loud.