The Statistical DJ

On Tuesday, streaming service Deezer announced a toggle to let listeners exclude AI-generated music from search results.¹ In a July report, the company noted that more than half of its daily uploads were generated by AI, averaging 90,000 tracks a day.² Bandcamp bans AI music,³ Spotify labels it,⁴ neither says what, exactly, is wrong with it. Perhaps the more interesting question isn’t who – or what – made it, but whether anyone will answer for it.

»The distinction isn’t whether a machine was involved, or even whether a human selected every note. It’s where the creative decisions reside, and whether anyone is prepared to stand behind them«

For a week, I’ve listened incessantly to Cameron Winter’s recent release, Live at Carnegie Hall.⁵ Paul Thomas Anderson is turning the performance into an upcoming concert film. There’s something deeply human about the setting: a sold-out concert hall, a grand piano, a 23-year-old musician alone on stage. Winter sings lyrics that seem to have arrived from somewhere far outside the statistical average:

You were born to break
my big hairy football arms
Like clean windows
kill the birds

In »The Rolling Stones«, there’s a moment that breaks me every time. Tissues at hand, I listen to a stranger articulate something I recognise but can’t quite name.

But Winter’s voice and piano are amplified. The performance reaches me through technology, just as all recorded music does. The history of music is, in part, a series of arguments about which technologies count as legitimate. Recordings severed music from the moment of its performance. Drum machines displaced drummers; samplers turned existing recordings into new compositions; Auto-Tune made the correction of a voice part of its aesthetic. Each new technology prompted a question about authenticity, and each time the boundary moved.⁶

AI is different, though perhaps not in the way its opponents suppose. A musician using AI to refine a composition is making different decisions from someone generating thousands of tracks and uploading them on autopilot. The distinction isn’t whether a machine was involved, or even whether a human selected every note. It’s where the creative decisions reside, and whether anyone is prepared to stand behind them.

Artist Sol LeWitt put the principle neatly: »The idea becomes a machine that makes the art«.⁷ Brian Eno’s generative music takes this seriously. The artist establishes a system of rules and constraints, then lets it produce music. Eno didn’t choose every note; often he didn’t even hear the result in advance.⁸ The absence of moment-to-moment intervention doesn’t make the work authorless. The decisions have moved upstream, into the design of the system itself. The idea makes the art.

But what happens when the cost of generating another ten thousand variations approaches zero? Someone still shaped the conditions under which the music emerges. But who chooses which results deserve to be heard? A system might become a kind of statistical DJ, producing endless variations and using criteria no listener can see to select among them. We’ve moved past the question of whether some human made the music to how much responsibility they took for that music.

Winter’s performance offers a possible answer. Playing and singing in real time involves countless tiny corrections: breath, pitch, timing, touch. Each is a risky decision, with no guarantee that the next moment will work. The performance is the accumulation of those decisions, made in front of thousands of people. I am in no way ruling out that a fully AI-generated song might make me cry. But hearing Winter live includes something that the recording can’t fully preserve: the knowledge that at a particular moment a particular person did a particular thing, and that it could have gone very wrong.

Perhaps that’s where music is heading. Recording made the musical object reproducible, pushing musicians towards live performance as a source of value. Maybe generative AI will intensify that movement. When music can be produced in unlimited quantities at negligible cost, the scarce thing might be the performance: the person who can actually play, sing, and risk something in real time.

Deezer’s toggle is a crude proxy for a much better question. »Was AI involved?« is relatively easy to answer. »Did anyone actually make a decision here?« is harder. A system can embody years of artistic judgement, or produce ten thousand variations that nobody has bothered to listen to. The label can’t tell us which. We’re trying to distinguish music that someone stands behind from music that merely exists. And no toggle can make that distinction for us.

References