The real question is not who created it, but what it cost

A musician once observed that the track he had spent six months writing and the one a machine had just produced in twelve seconds landed in the same folder, in the same format, under the same icon. Nothing in the file told one from the other.

The remark sounds anecdotal. It nonetheless poses the one question that will remain once production costs nothing: not what a thing is worth, but what it cost to make — and how anyone would know.

What the industry has already put in place

In July 2026 an unusually broad coalition — the international federation of the phonographic industry, its American counterpart, the independent associations A2IM, WIN and IMPALA, the Recording Academy behind the Grammys, the performers’ union SAG-AFTRA and the Human Artistry Campaign — presented two labels for recording metadata (Info Soir).

• AI-Generated: all or the essential part of the creative elements comes from an AI — a track produced from a text prompt, or with generated vocals or main instrumental parts.
• AI-Assisted: the creation remains fundamentally human, AI having contributed only to specific elements such as mixing or effects, with main vocals and instruments remaining human.

The distinction is finer than a mere warning, and that is what makes it useful: it accepts that the tool may play a part without the work ceasing to be human. One qualification is needed, however — this is an announced framework whose adoption is described as gradual and strictly voluntary. Nothing today obliges a platform to display it.

The standard underneath

The technical mechanism is older and reaches well beyond music. C2PA — the Coalition for Content Provenance and Authenticity, founded by Adobe, Arm, Intel, Microsoft and Truepic among others — defines a cryptographically signed manifest attached to the file (Content Authenticity Initiative).

The manifest can record the creator’s identity, the timestamp, the editing tools and processes used, the history of successive modifications, the other content used as ingredients, and a disclosure of generative AI. The signature guarantees the whole has not been altered. Newsrooms such as the BBC, Reuters and the Associated Press use it, as do camera makers — Leica, Nikon, Canon.

Note what that list contains and what it does not. It records events — a tool used, a modification, a date — because an event can be observed. It records no duration of attention, because a duration of attention leaves no trace in a file.

Its promoters are explicit about the limit: this information does not say whether content is false. It provides context, attribution, transparency. In other words, the standard answers “where does this come from” — it makes no claim to answer “what is this worth”, still less “what did this cost”.

The other proof, the missing one

The musician’s intuition went further than a label. What it calls for is not a proof of authorship but a proof of effort: a format that would record time spent, abandoned versions, the conditions of creation — the trace of the work rather than of the hand.

The idea appeals immediately, because it restores information the file has lost and that the listener once obtained otherwise: through price, scarcity, slowness, the story. All those indirect proofs rested on friction. They disappear with it.

Effort was never measured directly. It was inferred from what it made difficult.

Why it probably would not hold

Three objections, of which the last is the most awkward.

• It can be faked. Anything declared can be simulated. A proof of effort measured by the very machine used to create would be circumvented the day it was worth something on the market.
• It is intrusive. Recording working time, context, still more physiological signals, turns creation into a monitored space. The price is considerable, and it falls on the creator, not the machine.
• It presupposes what it ought to demonstrate. That effort makes value is not obvious at all. Major works have been written in a night, mediocre ones over ten years. Rewarding the toil confuses cost with result — the exact error of any badly framed set of accounts.

These objections do not disqualify the question; they disqualify the technical answer. What stands is the starting observation: we have lost a signal, and we have not decided what to replace it with.

What friction did without being asked

The first part of this series argued that friction — slow learning, error, checking — was the price of competence, and that removing it everywhere amounted to taking on an invisible debt. The question of proof is its outward extension.

Friction did not only produce competence. It also produced information about that competence, readable from the outside and without instruments. Difficult work was recognisable by its rarity. In removing the difficulty, we did not merely change how things are produced: we removed the signal that allowed one to tell them apart.

An open question remains, and Maeliom does not claim to settle it. Will tomorrow’s friction be returned by the machine, in the form of metadata, labels and signatures? Or will humans have to reintroduce it deliberately, in their organisations and in their work, by deciding what they go on doing themselves when they could be spared it?

The first answer is being written by the industry. The second belongs to nobody but those who lead — and that is Maeliom Consulting’s ground.

Common questions