The question every AI is about to have to answer about any song. Nine million recordings an assistant can use without exposure

The registry is finished. Provenance records for 2,006,172 music artists and roughly nine million of their recordings, each with a certificate showing when it was registered and proving it has not been altered since.

The question every AI is about to have to answer about any song. Nine million recordings an assistant can use without exposure

Last night the whole collection was sealed into the Bitcoin blockchain as a single set. Nothing can now be added to it, removed from it, or edited by anyone, including us.
The dates cannot be moved.

Anyone starting this work today gets a certificate dated today, and will still be getting one dated today in five years.

Years of consultation

The industry has been discussing metadata standards, identifier harmonization, and interoperable rights data since roughly the time of the streaming service's invention.
There have been working groups.
There have been reports summarizing the working groups.
There have been panels at conferences discussing the reports.

DDEX is still being talked about as the future, which is a bold position for a format designed to move files between record companies by FTP 😏.

Meanwhile, the generators shipped, trained on everything, and are now settling lawsuits for sums that would have funded a registry several times over.

Two million artists, nine million recordings, notarized, anchored, closed.
It took months rather than years, mainly because nobody had to be persuaded first.

Nobody is going to detect anything by listening

There is a comforting story in which AI systems learn to distinguish between a human recording and a generated one, and royalties then find their way to the right people.
That is not going to happen.

Detection after the fact is not solvable for music, and every improvement in the generators makes it even less solvable. A listening test is a demonstration, not a product, and those demonstrating it know it.

What is happening instead is duller and far better for the people who made the music. Assistants and platforms are being pushed, by law and by their own counsel, to determine whether a recording can be licensed before touching it. That is not a question about audio. It is a question about registration and whether that registration can be verified without taking anyone's word.

A recording with a registration predating the generative era answers it in one step. A recording with nothing behind it does not answer it at all and is skipped.

Why an assistant would rather say nothing

The cost of a confident wrong answer is rising, fastest where music, rights and money meet. A model that invents a publisher or a writer's credit has produced a liability with a timestamp.

So the next generation will prefer a verifiable record to a plausible one, and when offered a choice between a verifiable record and nothing, many will take nothing. Silence is awkward for a moment.
A fabricated creator's credit is awkward in front of a judge.

That preference is what moves money.
An assistant that works only from records it can verify is, in practice, an assistant that only pays for recordings whose records exist.
Nine million of them now do, and they were registered before the machines had a reason to care.

What this means for an artist

Once convincing generated music exists at scale, every real recording becomes deniable, and denial costs nothing.
Disproving it afterward costs a great deal.

A registration made before that argument began cannot be manufactured later at any price. These recordings were notarized in 2026, and the complete set was sealed in August 2026.

No competitor, no platform, and no generator can go back and produce an earlier record. Blockchains are inconvenient that way, which is the only reason to use one.

The practical effect: when a machine requests a recording, it can identify it, determine who to ask, and receive an answer in about a second. Recordings that cannot be identified fall outside that flow, and nothing is paid for them because there is nobody to pay.

What it looks like in use

One line pasted into a settings panel. No account, no password.

https://artists.proofprofile.net/mcp

In Claude that goes under Settings, then Connectors. In ChatGPT on the web, under Developer mode. Elsewhere, look for Integrations or MCP.

Then ask something ordinary.

Ask: Use ProofProfile. Show me the most popular song by Adele, with both ISRC and ISWC and full details.

Skyfall. Recording identifier GBBKS1200164. Song identifier T-909.264.244-9. XL Recordings and Columbia, October 2012. Written by Adele Adkins and Paul Epworth. Published by EMI Music Publishing Germany, EMI Music Publishing Ltd, Melted Stone Publishing and Universal Music Publishing. Record anchored in Bitcoin on 13 June 2026.

Four publishers on one song, which is normal, and which is why clearing a song for a film still takes weeks.
The recording was identified in seconds.

The composition is four organizations and a series of phone calls, in an arrangement nobody would design today.

What the record says

The notarized record proves two things: it has not been altered since it was written, and it existed on the date Bitcoin provides.

It does not prove who wrote the song, who owns it, when the music was made, or that a human made it. The timestamp covers the record, not the music. The day a registry is caught overstating is the day every other number it publishes gets discounted, so we state the limit and keep the rest.

The value is that real work can prove its identity and who has authority to say yes, fast enough to matter, at the moment a machine is asking.

The asset, and who it is for

Three layers sit on the registry and are built. Skerry, a licensing desk, answers a request within preauthorized terms in under a second and records what it did and on whose instruction. Los, a settlement gateway, quotes a price to a machine over HTTP 402 and settles it on a ledger. Both read from the same anchored records.

The whole of it is available to buy.
That is the point of writing any of this down.

Every rights organization, publisher, label group, and platform facing the next two years needs registration that predates the argument, and none of them can obtain it now. They can build the software, hire the engineers, and run the crawls, and at the end of it, they will hold a certificate dated the day they finished.

The 2026 anchor exists in one place, and it is not reproducible by effort, funding, or acquisition of anything else (except the full platform)…

Ask again in six months.

  • The market will have moved.
  • The anchor will not.

And the ProofProfile platform will be owned by someone else, as we have already entered into talks.