Cultural Essay

AI Can Write the Song Now: so what is left for the voice behind it?

Generative tools can produce a finished, radio-ready track in the time it takes to type a prompt. This is a cultural read on what survives that abundance — not a market forecast, but a look at what listeners still pay for once the recording itself is free.

Commissioned by · Marlon Bell · Media & Narrative

1. A voice memo at 2am

A songwriter finishes a chorus at two in the morning, hums it badly into her phone so she won't forget the melody, and uploads the same eight bars to a generative tool an hour later to hear it sung back in a stranger's borrowed voice, mixed and mastered before she has made coffee. Both files exist now. Only one of them cost her anything.

That small hesitation — approving a song rather than making one — is the feeling this essay is chasing. It is not about whether the tool is good. It is about what happens to the idea of a maker once the making is optional.

2. What the recording used to hold

For a century, a recording did two jobs at once: it captured a performance and it proved one had happened. Walter Benjamin called the thing lost in copying a work's aura — the presence of the original in time and place, the sense that a specific person, at a specific moment, made this exact sound. Recorded music always cheated that definition a little, but it kept a residue of it: a take, a room, a body.

Once a track can be produced without a performance behind it at all, the aura has nothing left to be a residue of. What listeners respond to next has to be built somewhere else.

3. The industry moves first

Culture is still arguing about what a machine-made song means. The industry, less patient, has already started pricing the argument. Through late 2025 and into 2026, the response ran on two tracks at once: settle with the model makers, then start labelling the output so someone can tell the difference.

A licence answers who gets paid. A badge answers who made it. Neither answers the harder question underneath both: whether listeners will actually go looking for the badge, or whether they will simply keep listening.

4. Wonder and fatigue, at once

Audiences are not behaving like one audience. The same feeds that reward AI-made spectacle also carry a countercurrent of exhaustion with it, and both readings are true of the same person on different days. A movement built around human-only art drew tens of millions of visits even as generated work saturated every platform it appeared on, according to reporting cited on the phenomenon. Fatigue and appetite are not opposite audiences — they overlap in the same listener.

That overlap is where the real market sits, not at either edge of it. A Luminate finding reported by Billboard put a number on the fatigue side: 42% of listeners said they were less interested in a song once they learned it was made with generative AI. Interest, in other words, is conditional on disclosure — which makes disclosure the thing worth watching, not the technology.

5. This has happened before

The anxiety that a machine might hollow out music is old. Synthesisers, drum machines, Auto-Tune and sampling all drew the same charge — that the human hand was being edited out of the record — and each time, the frightening tool eventually became a genre's signature sound rather than its replacement. Live performance survived recorded music for the same reason: Philip Auslander's account of liveness argues that a live event's value comes precisely from what mediatised culture has already made scarce, not from being technically unreproducible.

Every previous scare ended with the human role narrowing, not disappearing — curator, performer, brand, rather than sole producer of every sound. The question this time is whether that narrowed role still pays enough people to live on.

6. The turn: provenance is a claim, not a fact

The comfortable story is that a verified-human badge solves the trust problem — proof restored, listeners reassured. The uncomfortable turn is that provenance labels are themselves unverified claims sitting on top of a system built to be indistinguishable from the thing it labels. A badge only holds if the platform checking it is more rigorous than the tool trying to pass through it, and nothing in the current settlements guarantees that.

A label is only as trustworthy as the incentive of whoever applies it. That is not a technical problem the platforms can patch once; it is an ongoing one they have just agreed to police.

7. What can't be copied — yet

If the recording is no longer the scarce thing, the question becomes which parts of a musical life resist copying at all. Some of this is genuinely durable: a body on a stage, a story an audience already believes, a name with history attached to it. Some of it is durable only until someone builds a convincing synthetic version of it too — and several of those versions are already in beta.

None of these are safe forever. They are simply the things that currently take longer to fake convincingly than a chorus does — which is a temporary advantage, not a permanent one.

8. Three ways this settles

It is tempting to assume disclosure wins by default — that once labelling is reliable, listeners will sort themselves into human-preferring and AI-indifferent camps and the market will simply reflect that split. The counterargument is that most listening is not deliberate enough to sustain that sorting; people choose songs, not certificates, and the 42% who said they cared less once told may say something different from what they actually do at the next scroll.

Both outcomes are plausible from the same evidence. What decides between them is not the technology but whether caring about authorship turns out to be a value people hold, or one they only report holding when asked.

9. The chorus, revisited

The songwriter from the opening scene still has both files — the rough voice memo and the polished synthetic version. She has not deleted either. What has changed is not the sound of the song but her sense of what she is now being asked to supply: not a finished recording, which the machine can produce on demand, but a reason to believe a person stood behind it.

The open question is not whether audiences will notice the difference between a human and a machine. It is whether, once noticing takes effort, enough of them will still bother to look.

Grounded in Folka’s corpus: The Synthetic Sublime · The Model Mirror · Machine-Made Music

How we read this

This read draws on cited reporting on AI-music licensing, platform labelling and audience sentiment, set against long-standing media-studies concepts — Walter Benjamin's aura, Philip Auslander's liveness, parasocial attachment — to interpret what the data shows about authorship and belief. Figures are as reported by the outlets cited; the interpretation of what they mean for artists is Folka's own. The one number that would most change this read: a longitudinal study tracking whether audience preference for human-made music holds once disclosure becomes reliable and routine, rather than a one-off survey snapshot.