I Fed My Own Unreleased Song Into AI. Here Is What Came Back.
By Written by Peter McVeigh
It started with my own playlists. I began noticing tracks that felt slightly off, too clean in a way that didn't sit right, and the trail led back to AI. So I ran an experiment on my own unreleased song to see what's actually being gained, and lost.
Three Versions, One Song
It started with my own playlists. I began noticing tracks that felt slightly off, too clean in a way that didn't sit right, and when I dug into who had made them, the trail led back to AI. Then I spoke to a few friends with major publishing deals, people whose job is to sit in professional writing rooms every week, and they told me the same thing independently: Suno is in the room now. Not as a novelty. As a tool people actually use while they write.
Most of what gets written about AI and music is abstract. Will it replace songwriters. Will it end the industry. Is it art. I didn't want to write another one of those pieces, because none of them actually tell you anything you can use. So I ran an experiment on my own material instead, something specific enough to learn from rather than argue about.
Before you read any further, have a listen to all three versions. Everything below will make more sense with these in your ears first.
https://www.youtube.com/watch?v=2yO6hsQkoe8
https://www.youtube.com/watch?v=h9Agm0aehYQ
https://www.youtube.com/watch?v=MeJ5gg0vOTo
Why I Used My Own Song
I needed something where the copyright question wasn't muddy. So I picked a song I wrote entirely myself: the melody, the lyrics, the instrumentation, all of it. That way, whatever Suno did to it afterwards, I knew exactly what was mine going in, and I could see clearly what wasn't mine coming out.
I should be honest about my starting point too. I went into this with a preconceived opinion, and I still hold it: I believed a demo I made in Logic as a young, inexperienced producer would have more life in it than anything AI could generate from it. I wasn't trying to hide that bias. I was trying to test it properly, on a song where I had nothing to prove and nothing to lose either way.
Meet COLD
The song is called COLD. I wrote it seventeen years ago and never released it. It's a demo in the true sense: home-produced, unpolished, made by someone who was still working out what he was doing. That's exactly why it was the right song for this. There was no professional mix to protect, no reputation riding on the original, just a raw, honest piece of work I could hand over and watch what happened.
https://www.youtube.com/watch?v=2yO6hsQkoe8
What Suno Actually Does With a Song
I uploaded the original demo and let Suno analyse it and generate its own synopsis of what the song was doing, structurally and stylistically. From there, the actual work is in the prompting. I went through the description it gave me and started pushing it in different directions: EDM influence, Pop Punk, Pop Rock, seeing what each pull did to the song underneath.
"This is a male-vocal EDM Synth Leads, Euphoric dance music song in a minor key, featuring a driving rhythm section of drums and bass, distorted electric guitars, and a clean electric guitar."
That's it. One paragraph of description, fed into a tool that already had my original song to work from. That produced around sixteen versions. I picked the best one from the batch, which for me was the EDM-leaning version. That's worth knowing if you've never used one of these tools: you're not getting one AI opinion of your song, you're getting a wide spread of them, and the actual work shifts from writing to choosing.
That's worth sitting with for a second, because I think it's an argument nobody's making properly. Taste making is a skill. Knowing what a hit sounds like, being able to sit in a room with sixteen versions of the same idea and pick the one that actually works, that's not nothing. A&R has always been that job, in one form or another: someone with a good ear listening to a pile of options and knowing which one lands. Suno doesn't replace that ear. It just hands it sixteen times more material to use it on. Whether that's still "songwriting" in the way we've always meant it is a fair question. But it's not nothing either. It's just listening out for what you think actually sounds best, and being right about it often enough that people trust your judgement.
Taking It Further
I paid for a Suno Pro account so I could get underneath the finished track. Pro gives you the MIDI files and the audio stems, not just the final bounced version, which meant I could pull the arrangement into my own session and try something more interesting than just accepting what the AI handed me.
That's the actual session in Logic, shown above. You can see the stems Suno gave me broken out across separate tracks, drums, bass, guitars, synth, the outro chorus, backing vocals, sitting there exactly like any other multitrack session I'd build from scratch. What's missing is a lead vocal, because Suno's version had generated its own AI vocal performance and I stripped that back out. I muted it and recorded my own vocal over the top instead, singing the same lyric, in the phrasing and energy the AI arrangement had settled into.
That's the bit I actually think matters most out of the whole experiment. I haven't spent much time tuning or finishing that vocal yet, it's rough, but it's the most honest version of the question I actually wanted answered: what happens when you don't treat this as human versus AI, but as human and AI on the same track. The arrangement is entirely AI-generated. The performance on top of it is entirely mine. Neither half is pretending to be the other.
The Sonic Identity Problem
Here's the part that actually needs discussing properly, because it's more complicated than "who owns the song."
I own COLD. Every element that makes it a composition, the melody, the lyric, the chord structure, the instrumentation as I wrote it, that's mine, and nothing Suno did changes that. But the moment I asked it to reimagine the track as EDM or Pop Punk, it wasn't just rearranging my composition. It was applying a sonic identity, a whole vocabulary of production choices, textures and performance style, that came from somewhere else entirely: from whatever the model learned by listening to enormous amounts of other people's music. That identity isn't mine. It was never mine to license out in the first place, and I didn't ask for it to be attached to my song. It just was.
This isn't a hypothetical problem sitting in a courtroom somewhere unrelated to your career. It's already being fought over. Warner Music has settled with Suno and signed a licensing deal rather than keep litigating. Universal settled with Udio too, and the two of them are now building a joint, licensed AI platform due to launch in 2026 with opt-in artist compensation. But Sony hasn't settled with either company, and Universal and Sony's case against Suno is still active, with no trial date yet and the judge still to rule on summary judgment. In other words: the biggest companies in this industry haven't agreed among themselves what's fair use and what isn't. If they haven't worked it out, you should assume it isn't settled for you either.
If you're a manager and an artist you represent starts experimenting with these tools, or a label starts pitching you on AI-assisted writing sessions, this is the actual question to ask: whose sonic identity is on this track, and has anyone agreed that's allowed to be there.
So let me put it as plainly as I can. Copyright-wise, COLD is mine. It's all my hooks, my chord progressions, my melody, my lyric. So what's actually being taken here? It's someone else's sonic identity, some producer's, some artist's, some engineer's fingerprint that the model learned from listening to their work, without them agreeing to hand it over. Is that the future of music? I think part of it genuinely is.
Do I think bands, artists and writers are going to disappear? They've been disappearing for a while now, that's not new, and it's not really an AI story. What I think actually happens if people lean into these tools more and more is that less money flows into the industry overall. That's the real damage. Less money means fewer people can make this their profession, full stop. That's the bit that should worry anyone reading this who works in music for a living, not the existential "is it art" question.
I can see two ways this settles. Either a legal challenge forces a proper outcome, one where music that was scraped to train these models gets compensated through some kind of royalty, or the sites that won't play fair get shut down entirely. Right now we're watching both tracks run at once: Warner and Universal choosing to licence and get paid, Sony choosing to keep fighting in court.
But here's the part I don't think slows down either way. With how fast LLMs and the hardware running them are improving, I don't think there's any stopping this long term. Quick explanation if you haven't had to learn this yet: an LLM, a large language model, is a type of AI trained on enormous amounts of existing material so it can generate new material that follows the same patterns. It's the same underlying idea whether that material is text, images, or, in Suno's case, audio. Training a model like that takes huge computing power. But once it's trained, actually running it is a much smaller job, really just number-crunching. If a computer is powerful enough to hold the model and do that maths quickly, it can run the whole thing offline: no internet connection, no company's servers, no usage logs, nobody watching what you feed it.
That's the direction this is heading. Give it a few years and people will be able to feed their own computer a library of their favourite artists' music and generate their own version of Suno, entirely offline, on a high-powered machine they own outright.
Which is why I'd keep an eye on Apple specifically. They've been quiet in public on the AI front compared to everyone shouting about it, but from what I've been reading, their approach has been to build the on-device side rather than compete on chatbots: real investment in local, private, on-device AI processing, with recent Apple Silicon generations adding dedicated AI accelerators direct into the chip and unified memory that lets a Mac hold and run much bigger models locally than a typical PC graphics card can manage. If that keeps developing at the pace it's going, Apple's hardware, more than any single AI music company, might end up being what makes fully offline, fully personal AI music generation an ordinary thing people just have on their laptop.
What Human vs AI Actually Showed Me
I set up a YouTube channel called Human vs AI and put the results in front of real listeners, deliberately releasing the AI version under a different band name so it wouldn't get tangled up with searches for my own music.
The reaction was consistent. People preferred the original demo. They said it had more life in it, more feeling, even with all its rough edges. Here's what actually got left in the comments on the original:
"No question this one is far superior. Instruments sound way better and I actually believe that the lyrics being sung in this version came straight from an actual humans experiences." — @lapsemusic579
"The high d note that rings out a lot throughout the demo on the 3rd beats is a real centrepiece of the song for me, so honestly even for that arrangement difference alone I'd say I prefer the demo. I prefer the drums in the original as they feel more driving and unique, the AI drums sound more generic, although I do like some of the AI version's guitar 'ideas', and you could say the AI version's chord progressions provide more movement at times, in the choruses for example. I like the more frequent pauses in the original, as the dynamic contrast heightens the emotional impact for me and leaves some breathing room, at times the AI version feels like it's taking the whole wall of sound thing to another level. Again, I feel like for all of these videos I'm always gonna say I prefer the vocals in the original songs because they're more emotive and expressive, AI vocals feel static to me. I just feel that as big and 'exact' the AI version is, things can be too perfect, to the point they're not interesting, so they ironically become imperfect, to my ears at least." — @hollygallagher1745
Nobody disputed that the AI version was better produced, cleaner, more polished, more "finished" sounding in the way a modern release is expected to sound. But that's not the same as people preferring it. On the AI version itself:
https://www.youtube.com/watch?v=h9Agm0aehYQ
https://www.youtube.com/watch?v=MeJ5gg0vOTo
"This version sounds so much like the royalty free music that plays where I work lol." — @lapsemusic579
"I was blown away the first time I heard the original version because of the passion in it. This is good but it's way too clinical for me!" — @critter-cat
"It's not unpleasant, but knowing how good your original version is I feel this one is 'plastic'. The AI produced tracks I've heard are too precise, no humanity to them, but do people care enough these days?" — @critter-cat
That last question is the one that actually matters. "Too clinical," "too perfect," "plastic," that's the same complaint from three different people who don't know each other, and it lines up with the audio artifact problem I heard myself: the AI audio is full of small glitches and textures that don't belong, and I'm honestly not sure how usable it is in its raw state without a lot of cleanup work.
I'll say the obvious caveat myself: this was one song, judged by people who chose to watch a channel called Human vs AI, so they were never a neutral panel. But the pattern matched what I expected going in, and it matched what my friends in professional writing rooms are also finding: the tool is fast, the tool is often technically impressive, and the tool still isn't making the decision about whether something actually moves someone. That's still a human job. The open question in @critter-cat's comment is the one I don't have an answer to yet: do people care enough about that difference to keep paying for it.
So, Are We Safe?
I have no interest in using these tools. Writing music is like working on a puzzle for me. I'm not looking to get a book or a YouTube tutorial on how to do it. I want the frustration. I want the fact that I maybe didn't quite get it right. Writing music, for me, is about learning and failing and trying harder next time. Using Suno just feels, to me, like pressing next on my iPod or Spotify and stopping when something sounds familiar. There's nothing challenging about using it for me, and at this point I don't see it, or tools like it, as part of my career.
Musicians and songwriters aren't obsolete. Nothing in this experiment told me that. What it told me is that pretending these tools aren't already in the room doesn't help anyone. They are. Major label money is actively going into legitimising them right now, not banning them.
What that means practically, if you manage artists or you are one: know what these tools can actually do before you have an opinion about them, because your opinion is more useful to your artists if it's informed. Understand that owning your composition is not the same as owning everything a tool does to it once you feed it in. And watch the licensing deals happening between labels and platforms like Suno and Udio closely, because whatever terms get set there will very likely become the terms everyone else works under eventually.
I'd genuinely like to know what you hear. Post it in the forum, tell me which version actually moves you, and tell me if what's happening in your own writing rooms matches what I'm describing here. This is going to keep changing fast, and I'd rather figure it out with people actually doing the work than guess at it alone.
Peter McVeigh is the founder of Music Manager Resource.