Where There's a Right Answer

HAUSVERBOT put a clip of Marcel Dettmann online in the same week Bandcamp banned AI-generated music from its platform. He is asked about the fear that AI will make DJing irrelevant — that the robots will simply play the recommendations.

He is unbothered by that framing, and points somewhere else entirely:

Optimizing stuff, it's the killer.

His argument, paraphrasing: use it to be creative or to optimise whatever you already did, and you get lost, because the possibilities are endless. You will never be happy. You are looking for perfection. He is careful not to call it the devil — it is a nice tool, and it is genuinely helpful here and there. Just not for the creative part.

The danger is not the machine

The interesting move is that he declines the science-fiction version of the question. Nobody serious believes a model is about to take a Berghain booking.

What he describes is psychological and much more likely: a producer with infinite versions available, optimising forever, converging on nothing. Before these tools, the number of variations you could physically make was capped by time. That cap was doing hidden work — it forced decisions. Remove it and "finished" stops being a state you arrive at and becomes a choice you have to make against infinite unexplored alternatives.

Perfection was always unavailable. The difference is that it now looks reachable.

The distinction that saves the tool

Here is the rule I have settled on, and I think it resolves most of the argument.

Some questions in music production have right answers, and they are checkable. Is there a resonant build-up at 240Hz that will make the mix muddy on a big system? Is the low end mono below 120? Is the master hitting its target loudness without pumping? Are two elements masking each other? These are measurements. A tool that finds them faster than your ears do is a better ruler, not a co-author, and refusing it is superstition.

Other questions have no right answer at all. Should this break at 3:20 or run another ninety seconds? Should the hi-hat sit slightly wrong? Is this track about something? Optimisation needs a target, and when there is no correct answer, the only target available to a model is the average of everything it has already been shown. Which is precisely how you get the frictionless, unskippable music I described in what the metrics measure — pleasant, legible, and gone.

How we use it here

At Hybrid Flight the line is drawn exactly there, and it is not a moral position, just a practical one.

AI-assisted tools are welcome for the technical layer. Spectrum analysis to see what my ears have stopped noticing at two in the morning. Mixdown checks. A sanity pass on the master. All of that is fair, and it makes the record better in ways that can be verified rather than argued about.

What those tools deliver is a technically correct foundation — and that is all they are permitted to deliver. A technically correct foundation is not a track. It is the floor the track stands on.

The one grey area worth naming honestly: using a generator or a random modulation to produce a starting point is fine, and I have no problem with it. A generated pattern is raw material like a field recording or a happy accident with a broken pedal. The rule is only that it never survives one to one. Something generated is a suggestion, and taking a suggestion without changing it is how you end up releasing the average of a training set with your name on it. Your stamp goes on it or it does not go out.

The other half of the argument

It would be easy to leave this as a defensive piece, so it is worth reading it against something less cautious.

Matt Bluemink's The Future is Not Lost: On Music, Technology, and the Creation of New Worlds (Becoming Press, 2026) argues more or less the opposite mood. It's a set of essays on music, technology, architecture and philosophy, drawing on SOPHIE, Arca and Iglooghost, and its case is that the future has not evaporated — that it still gets imagined, and that music and technology are where that imagining happens.

That is not incompatible with Dettmann. SOPHIE's work was radically technological and sounded like nothing else, which is the whole point: the technology was in service of a decision nobody else would have made. The failure mode is not using new tools. It is letting the tool make the choice, because tools trained on the past can only ever propose the past, slightly rearranged.

Dettmann is warning about surrender. Bluemink is describing what happens when there is no surrender. Both can be true, and if you only read the pessimistic half you will end up making conservative records for good reasons, which is still making conservative records.

What Bandcamp's ban actually says

The context for the clip: in January 2026 Bandcamp became the first major music platform to ban AI-generated music outright — nothing generated wholly or in substantial part by AI, no using it to imitate specific artists, and no scraping its catalogue to train models. Their reasoning was that listeners should be able to trust that what they find there was made by people. Subvert, the co-op I wrote about in the equity nobody shared, took the same position from the start.

A ban is a blunt instrument and enforcement will be genuinely difficult — the line between a mastering assistant and a generative model is not obvious from the outside, and a determined uploader will not be stopped by a policy. But the signal matters more than the enforcement. It is a marketplace stating that the thing it sells is human authorship, at a moment when the largest platforms are quietly discovering that they can sell something cheaper.

Which returns to the only rule I would hand anybody. Let the machine measure. Do not let it decide.

Source: Marcel Dettmann in conversation with HAUSVERBOT, clip posted 22 January 2026; the full interview is on their YouTube channel. The quoted line follows that clip. Bandcamp policy details are from its January 2026 guideline update as reported at the time.