Market Commentary

The Human-Made Label: why a mark of authorship is easier to fake than to prove

Platforms are racing to tag what AI made and what it didn't, but the labels disagree on what they are for. This note reads the labelling moves of 2026 against the evidence on whether a human-made tag changes anyone's behaviour, and lays out where the market is actually heading.

By Trust desk × Folka

1. A claim that needs proving

On 14 October, YouTube Music began tagging tracks Fully, Partly or No Gen AI, built on partner declarations plus its own detection. The same week, The Document Foundation shipped LibreOffice 26.8 under the banner 'no AI is now a feature.' Both moves sit on the same premise: that 'made by a person' is now something a product has to state, not something a listener or reader assumes by default. That premise did not exist five years ago. It exists now because the alternative — generated work indistinguishable from human work, moving through the same pipes — has become common enough that silence reads as ambiguity.

That gap — between what gets marked and what gets claimed — is the seam the rest of this market runs along. The three platforms that have moved first have not converged on what the mark should do once it exists.

2. Three platforms, three bets

Deezer, Spotify and YouTube Music all now label AI involvement, and all three do something different with the label. Deezer excludes AI-flagged tracks from recommendation outright. Spotify keeps labelled 'AI Persona' artists out of recommendation by default, a narrower cut aimed at synthetic performers rather than AI-assisted human ones. YouTube Music discloses and keeps paying: AI involvement, including fully AI-made tracks, does not by itself cost a track its monetisation, and the platform has not said whether the label touches ranking at all. A creative team's exposure to any one of these rules depends entirely on which platform's bet it is standing on, not on whether a label exists.

Filtering treats AI involvement itself as the risk to manage. Disclosure treats the listener's informed choice as the end of the platform's job. Those are not two versions of the same policy; they are two different theories of what a label is for, and nothing forces them to converge.

3. The supply is a flood; the listening is a trickle

Deezer's own figures complicate the case for urgency. The company said more than half its daily uploads — roughly 90,000 tracks a day in June — were AI-made. But AI tracks account for only 1 to 3% of its streams, and Deezer separately says up to 85% of the streams those tracks did get in 2025 were fraudulent. Read together, the numbers describe a supply-side flood that barely touches real listening, propped up largely by streams nobody actually heard.

That math reframes the labelling race as a fraud-and-royalty-pool problem as much as a trust problem — the two motives happen to point the same way, which is part of why platforms moved on this so fast in 2026.

4. The premium nobody notices

Set the platform mechanics aside and ask whether human-made work actually performs better with audiences. Ipsos, working with Syracuse's Newhouse School, tested 20 ads across 3,000 US consumers in a study published 18 May. Human-made ads outperformed AI-made ads by roughly 14% on short-term sales response and 17% on long-term brand-equity measures. The twist is in who could tell the difference: only 25% of people shown the AI ads were confident they were AI-made, even as 79% said they want companies to disclose AI use.

Read the other way from Ipsos's own headline, the finding is stranger than 'AI ads are good enough.' A real performance gap exists between human and AI-made work, and it operates below the level where people can consciously spot the difference. That undercuts the simplest case for a human-made badge: if the premium already works on people who cannot identify the source, telling them explicitly may add little, or may even interrupt a response that was working unlabelled.

5. Why the label can subtract

There is a second, sharper asymmetry: an AI label does not just fail to add a premium, it can actively cost one. Wu at Stanford and Holmes at Reed, writing in Cognitive Research: Principles and Implications, found listeners engaged less with music labelled AI-made, even when a person had actually made it. That effect has a known name and a two-decade history. Kruger and colleagues described an 'effort heuristic' in 2004: people rate a poem or painting more highly when told it took longer to make, even when the work itself is identical.

The replications support part of the original effect, not all of it, so this reads as a real tendency rather than a settled law. But it explains both puzzles at once. 'Human-made' signals effort without having to say so; 'AI-made' signals the absence of effort, whether or not a person actually did the work. It also implies a sharper design question than a binary badge answers: a label that showed process — hours, drafts, hands involved — might carry more than a yes/no AI tag ever could.

6. What 'no AI' is actually selling

LibreOffice's 26.8 release, shipped 26 August with 'no generative AI features,' is the cleanest test of what a human-only claim is for. The Document Foundation's own framing — 'Yes, no AI is now a feature' — leaned on user control and the absence of telemetry, not on craft or authorship. Nobody buys open-source office software for the emotional authenticity of its formatting engine.

That ambiguity is a problem for anyone trying to build a market around the human-made label, because a buyer cannot tell, from the badge alone, whether they are purchasing craft, privacy, or simply a product's refusal to automate.

7. What the rules agree on

Russia has said it will require AI labelling from March 2027. That announcement reads as a first mover only if the rest of the field is ignored. China has required visible, machine-readable labels on AI-generated content since September 2025. The EU AI Act's transparency duties, covering AI-generated content disclosure, have applied since August 2026. What the three regimes share matters more than their dates: every one of them marks the machine's output. None certifies the human.

A watermark can prove a model touched the work. No regulator anywhere has proposed a mechanism to prove a model did not — that absence has to be declared and then audited, which is exactly the gap several private certification marks are now trying to fill, with no shared standard between them.

8. Who absorbs a wrong call

A labelling or detection system that runs at platform scale will misclassify some real work. A human track wrongly flagged as AI-made, or a hybrid piece that falls between a platform's declared tiers, loses recommendation reach under a filtering regime like Deezer's or Spotify's, exposure under a disclosure regime like YouTube Music's, or simply credibility if a certification mark gets it wrong. Follow where that cost actually lands.

That last step is where the asymmetry sits. A major label or a well-represented artist can contest a wrong call; a small, independent creator usually cannot. Nobody — not Deezer, not YouTube, not Spotify — has published a false-positive rate. Until one does, the actual error cost of this infrastructure is unmeasured, and it is very unlikely to be distributed evenly.

9. Our view

We think the labelling race is being read as a trust fix when it is mostly a fraud and royalty-pool fix wearing trust's clothing. Deezer's own numbers make the point starkly: AI supply is enormous, AI listening is marginal, and most of the streams AI content did receive were fraudulent rather than real engagement. Filtering AI-flagged work out of recommendation protects the royalty pool from exactly that kind of gaming, and it would make commercial sense even if no listener cared about authorship at all. The fact that it also reads as consumer protection is useful cover, not the primary driver.

Two forks sit underneath this, and neither resolves cleanly. One is whether platforms converge on filtering, like Deezer and Spotify's AI-Persona rule, or disclosure-while-paying, like YouTube Music — the assumption to watch is that fraud losses, not listener sentiment, decide which model a platform picks, so a platform with a bigger fraud problem is more likely to filter. The other is whether certification marks consolidate into a shared standard or stay a crowded field of competing private badges — the assumption to watch is that a shared standard needs a regulator or a dominant platform to force it, and none has moved to do so yet. A credible disclosed false-positive rate from any major platform would be the clearest sign either fork is tipping. Open questions that stay with the reader, not with us: whether China's, the EU's and Russia's labelling regimes converge enough to make cross-border provenance claims interoperable; whether any platform publishes a false-positive rate for its AI detection; and whether a process-based label — showing hours or drafts rather than a yes/no tag — gets tested anywhere against the effort-heuristic finding that seems to predict it would work better.

10. What businesses are actually choosing between

Three different futures are visible in the moves already made, and none of the evidence available says which one wins. Each rests on an assumption that could break.

Each path changes what a creative team is actually being judged on — detectability, disclosure, or certification — and each carries a different kind of exposure. None of the evidence gathered here points firmly to one outcome over the others.

How we read this

This note draws on platform policy statements (YouTube Music, Deezer, Spotify, The Document Foundation), one peer-reviewed study (Wu & Holmes, Cognitive Research: Principles and Implications), one published market research study (Ipsos with Syracuse University's Newhouse School, 18 May 2026), Deezer's own disclosed fraud and upload figures, and the stated regulatory timelines for China, the EU and Russia. The effort-heuristic citation (Kruger et al., 2004) is foundational psychology with partial later replication, used as interpretive context rather than a measured figure for this market. The single number that would most change this read is a platform-published false-positive rate for AI-content detection — none of the sources reviewed here discloses one.

Grounded in Folka’s corpus: The Human Seal · Handmade, On Purpose · The Provenance Premium · The Filter for Slop