Provenance: Why "Who Made This, When" Suddenly Matters

For most of recorded music’s history, provenance took care of itself. Making a record required a studio, an engineer, a label, a pressing plant — a supply chain of witnesses. If a question ever arose about who made something and when, there were invoices, tape boxes with dates written in marker, union session sheets. Provenance was a byproduct of how expensive production was.

Then production got cheap, and the witnesses went away. A finished track now comes out of a bedroom with no paper trail beyond file-modified dates that change every time a project is copied. And in the last few years, generative models removed the final barrier: it no longer takes a human performance to produce a convincing recording — of a genre, of a style, of a specific artist’s voice. The result is a strange inversion. Music has never been easier to make, and a claim about who made a piece of music has never been harder to check.

What provenance actually is

Provenance is the art world’s word, and the art world’s version is instructive. A painting with provenance comes with a documented chain: painted by X in year Y, sold to Z, exhibited here, authenticated there. Each link is a checkable fact. The painting without provenance might be genuine — but every transaction involving it now carries the cost of that uncertainty, and that cost comes out of the seller’s price.

Musical provenance is the same structure: a set of checkable facts about a work’s origin. This exact file existed by this date. This person claimed it, publicly, at that time. These people were declared as contributors, in these shares, before any money was at stake. Not opinions, not vibes — facts with timestamps, checkable by a stranger without trusting you.

Why the AI era raises the stakes

Three separate pressures are converging on the same weak point.

Disputes are getting harder to referee. When anyone can generate a track in a given style, “I made this first” becomes both more common as a claim and less obvious as a fact. Platforms adjudicating takedowns, distributors screening uploads, and courts weighing infringement all face the same question — which of these parties can prove anything about when their file existed? — and mostly find that neither can. In that environment, the party who happens to hold a dated, signed, third-party record starts every argument several moves ahead.

Training is a new use of your work, negotiated without you. Whatever position you take on AI training — opt in, opt out, license it, fight it — every position requires the same substrate: a record of what you own and what you said about it. Consent and compensation systems for training data are being designed right now, by companies and by legislators. Those systems will key on machine-readable claims. A catalog whose ownership lives in your head is invisible to them; a catalog with public, signed, timestamped records is legible infrastructure those systems can actually respect.

Authenticity is becoming a market signal. As synthetic media floods every channel, “verifiably made by a person, on a date, by this artist” shifts from assumed to asserted — and assertions need evidence. Standards like C2PA on the capture side and provenance registries on the claims side are two halves of the same response. Listeners may never inspect a hash themselves. But platforms, licensees, and supervisors increasingly will, the way buyers check a car’s title without caring how title databases work.

What a good provenance record looks like

Not all records are equal. The useful ones share four properties, and it is worth knowing why each matters:

• Exactness. The record ties to a specific file via a cryptographic fingerprint, not to a title. “I made a song called Midnight” proves nothing; “I claimed the file with this hash” can be checked against any copy, byte for byte.
• Signature. The claim is cryptographically signed, so it attributes to a specific keyholder rather than to whoever typed a name into a form.
• Independent timestamp. The date comes from a system the claimant does not control — a public blockchain’s block time, in this registry’s case — so it cannot be backdated.
• Public checkability. Anyone can verify the record without asking permission or trusting the registry’s website. Evidence you must ask the other side to confirm is weak evidence.

Notice what is not on the list: any assertion that the claim is true. A provenance record proves that a claim was made — exactly, by a specific key, at a specific time. It cannot prove the claimant wrote the song. That limit is not a flaw; it is what makes the rest trustworthy. Systems that promise more than their mechanism can deliver end up delivering nothing.

The timestamp asymmetry

One property of timestamps deserves its own paragraph, because it drives the practical advice. A timestamp proves your claim existed no later than a date. It says nothing about earlier. This creates an asymmetry: registering early costs almost nothing and preserves the strongest possible bound, while registering late permanently caps what you can prove. You cannot go back and file 2024’s song with 2024’s date. Every month a finished work sits unregistered is provable priority you are giving away — to no one in particular, which is somehow worse.

What to actually do

The practice that falls out of all this is short. Fingerprint and register finished works when they are finished, from a wallet or key you actually control. Declare contributors at filing time, matching a written split agreement. Keep the exact registered files somewhere safe, because a fingerprint is only useful if you retain a file that matches it. And be precise about what your records mean when you point to them — a provenance record is evidence, not a verdict, and the people most worth convincing are exactly the people who know the difference.

The bet behind all of this is simple: in a world where anyone can generate anything, the scarce good is not content but accountable origin. That was always true in art. It is becoming true in music at machine speed. The artists who will navigate it best are not the ones with the strongest opinions about AI — they are the ones whose catalogs can answer, precisely and publicly, the oldest question in the business: who made this, and when?