AI music has been reviewed by many people with good headphones.
Journalists review it, producers dissect it, and tech writers run it through a benchmark and report back. All perfectly legitimate. The people who have to play it in a room full of strangers, who paid for their drinks and expect something in return, care less about whether AI music can be heard than whether it can be used. That is the question I ask in this new age: can it be worked with?
A DJ’s relationship with a record is different. I don’t just hear it; I use it. I cue it, loop it, pull the bass out of it, lean on it for a few minutes while something else fades in, and then I find out, within eight bars, whether I was right. I’ve had records that looked fine until I tried to ride them for a minute and a half, and others that only proved themselves when the next track came in. That’s the lens I want to apply to AI.
Instead of asking if it sounds like a song, I want to know if I can work with it.
The 10-second problem
It helps to start with how some of these models have been evaluated, because they’re judged in very small pieces. MusicLM’s listening tests, built on Google’s MusicCaps dataset, asked raters to compare pairs of 10-second samples. A larger study from ETH Zurich used 6,000 songs from 12 models and thousands of listener comparisons. Likewise, it restricted the listening comparisons to 10-second instrumental clips.
Ten seconds is a perfectly reasonable unit of measure if you want to know whether a model can create a convincing piece of music. It’s less useful if you want to know whether it can hold together for six minutes, survive a pitch change, or blend with another record through a club’s sound system. A club track has a lot more on its shoulders, so the test should match the use.
The blind test everyone quotes
A Deezer and Ipsos survey of 9,000 people found that 97% failed its three-track test of identifying fully AI-generated and human-made music. That finding has become part of the wider debate over identifying and labeling AI music, but it says little about how a track behaves in a mix.
If listeners miss the join, “can you spot the robot?” becomes a less useful question for a DJ. The one I’m asking is more practical: how much work do I have to do if I’m going to play this file to a dancefloor?
What a DJ is actually asking
Before a record gets into a set, it has to go through a fairly quiet form of vetting.
Does the grid hold from the first kick to the last, or does it wander off around the third minute?
Is there an intro I can use without guessing?
When I bring it in over the top, is there room for both, or does it fight the low end?
Can I loop it, filter it or pitch it and still like what I’m hearing?
And if the room turns on me halfway through, is there a way out?
In the booth, I know within a minute or two whether I’m fighting the blend or shaping it. I can hear it when the incoming track sits too hard against the outgoing one, and feel it when the blend stops opening up.
None of that is guaranteed by a genre tag. “128 BPM house track with a rolling bassline” describes a genre. It says nothing about thirty-two bars of usable drums at the start, a breakdown that earns its drop, or a second drop that does something different from the first.
This is where generated tracks can get interesting, and at times awkward. One can work at the first bar and feel unconvincing by the fourth minute. Phrase lengths can wander. The stereo image can sound huge on headphones and thinner once the low end is summed to mono. A drum pattern can look tidy on the waveform and still feel different in your hands, which you only notice when you try to ride it for a minute and a half.
Built for the single, not the blend
Another tendency is easy to overstate, but short, song-shaped output can sit awkwardly in a long blend. The ceiling isn’t the cause, as Suno says it can generate up to 8 minutes in one pass. Length alone doesn’t guarantee a usable arrangement, though. A short structure also resembles the music the tools sit alongside. Chartmetric found the average Spotify charting song in 2024 was about three minutes long. A quick hook, then home in time for the next track.
Whether that’s a problem depends on who’s in front of you. In a mobile DJ environment, a short intro and a front-loaded hook can be useful. Weddings, birthdays and parties are full of casual listeners and pop enthusiasts who want the instant gratification of familiarity, so quick edits often work well. A generated track can fit that pacing, even if it still has to earn its place alongside songs the crowd already knows.
Electronic dance culture also has a different tradition. House, techno, garage, and drum and bass all have a history of extended mixes, long intros, patient breakdowns and plenty of time spent blending two records. The mix was integral to the music and to the DJ’s skill.
Hand a DJ working in that tradition a three-minute track with the hook at the 30-second mark and you’ve given the blend very little to work with. The format can sit awkwardly against those habits.
One record can sound strong on its own and still leave the next one nowhere to sit. That is where AI music still has to prove itself: in the last transition as much as the first listen.
It’s fair to say this culture isn’t uniform, and some of it has drifted towards faster builds and more frequent drops. In sets built around quick edits and frequent drops, a short generated track may be a comfortable fit. The friction I’m interested in is at the other end, in the underground rooms where patience is part of the appeal and the long, slow turn of a proper blend is the whole event. That’s an observation from the booth, not a finding from a study.
Where a human steps back in
The picture changes once there’s something to edit. Parts of the industry are moving towards a more licensed footing: Suno’s November 2025 partnership with Warner Music Group was followed by the launch of V6 in September 2026, developed with industry partners including Warner, BMG and Believe. As platforms mature, generation can become less of a finished product and more of a raw material: a drum loop you might want to keep, a vocal phrase you might want to chop, a bassline you might want to replace.
That’s where the producer or DJ comes in. You pick the best eight bars, build the arrangement with a proper intro and outro, clean up the low end, and you’re closer to a track you’d be happy to play on a Saturday night.
If you skip all that, you can be left with an anonymous file that makes a great first impression and has no direction for mixing out. In the booth, the difference shows up fast: one version gives me a clean way in and out, and the other leaves me forcing the handover.
The ones already doing it live
Some have skipped the finished track entirely. Nao Tokui’s Emergent Rhythm set at MUTEK Japan used generative models during the performance itself, and the work was also screened in the NeurIPS Creative AI track. Dadabots, famous for their 24/7 neural metal streams, describe what they do as “prompt jockeying”: generating tracks on stage and mixing material they’ve never heard. They describe it as making DJing harder, since the tracks don’t exist yet. That’s pretty much the sum of working without a net.
This is the most promising route, I think.
A model that responds to the performer at the level of loops, bars and phrases suits a DJ far better than one that hands over a finished record and vanishes.
It keeps the person in the loop, and they’re the one who can see the room.
Who is responsible for the track?
Copyright and ownership questions still surround all of this. The US Copyright Office’s report on copyrightability draws a line between AI assisting human expression and AI determining it. Separately, on the question of training, the UK government backed away from its preferred text and data mining exception in March 2026, saying it “no longer has a preferred option”. Those are different legal questions. In the booth, there’s a more immediate kind of responsibility.
What a crowd gets from a DJ is judgement: why that record, for those people, at that time. A model can generate a file, but it doesn’t stand behind the decks watching the floor thin and deciding what to do. When a track dies, I have to read it, adjust and move on. The blame is mine, and I haven’t seen a platform offer to share it yet.
A brief for the next model
If I were to write the spec for the next generation of these tools, it wouldn’t be in the usual genre-tag format. I’d want an intro I can enter without staring at the waveform, and a tempo that stays put unless there’s a good reason to change it. I’d want the kick and bass kept apart so they can be treated as two separate problems, and a second drop that develops the first. I’d be as happy with a six or seven-minute piece as a three-minute one, and I’d want stems that behave like stems, with more than one way out.
That’s the brief: usable from first cue to final transition.
None of that asks for better taste. It asks the model to understand how music is used as well as how it sounds. The first system to get there will reach a DJ’s crate fastest, because it will give the DJ more room to respond to the floor. That’s the standard I want the tools to meet: a better ending as well as a better output.