Consented, artist-owned AI models that pay the people whose work made them. A free-to-use toolkit that lets a musician train an AI model on their own catalogue, locally, on hardware they already own.
A musician trains it on their own catalogue, on hardware they already own. The catalogue is theirs, so the consent question is answered before any training starts. Whatever the model makes carries attribution back to them.
Most of the response to AI in music has been open letters, consultations and lawsuits, all of it aimed at tools that were already built, on other people's work. A working alternative, built rather than argued for.
Free to use, and it runs on the musician's own machine. They train a model on their own catalogue, and nothing leaves the machine unless they choose to send it. The model belongs to them, with no platform claim and no exclusivity.
Working prototype now, public release in 2027.
A listener goes to one artist's space and prompts within that artist's sound, and hears what that artist's model makes for them. The artist sets the price and keeps the relationship with the listener.
In proof of concept, public release in 2027.
Before anything is trained, every track is checked: who holds the master, the song, the performances. Fingerprint, ISRC and agreement checks flag anything with an uncleared co-writer, sample or session player.
Each writer and performer is asked in writing, for this purpose only, with the right to withdraw. No default opt-ins. Those who say no are left out and lose nothing. A voice is never trained without separate consent.
The model trains on the artist's own machine. The catalogue goes to no third-party AI company. A usable model needs as few as eight tracks; twenty or more is better.
The model is checked for memorisation. A model that reproduces a training track is not released.
The model sits in the Listening Space. The owner sets what can be prompted, who gets in and what a session costs. By default nothing can be downloaded or released.
Each output is watermarked, stamped with provenance data and logged with the share each training track contributed. Statements go out each quarter, split by the attribution record.
Take a piece of music made by an artist's own model. Strip the metadata, re-encode it as an MP3, and put it through everything that normally destroys provenance.
The attribution is still there, and it still names the artist.
Under MP3 at 128 kbit/s, a 6 dB level change or a 10% trim, no share moved by more than 0.05.
Estimated share of one output attributable to each track in the training catalogue.
How the writer share of that output divides, by the attribution record.
The split itself was tested against a retraining ground truth, with the bars set in writing before any result was scored:
These are estimates with a stated error, and the tests that failed are reported alongside the ones that passed. The working paper behind these numbers, with the method and the full results, is published here.
Every piece of audio made with Khaos carries an inaudible watermark and a signed record saying it is AI-generated. The verifier is free to use, and answers with one of five verdicts.
Confirmed
The audio matches a released Khaos piece.
These promises are written into both companies' articles of association.
A non-profit company limited by guarantee. It owns the Khaos code and the attribution standard, and holds a guardian share in the operating company with a veto over any sale, change of purpose, or revision of the fee caps.
The operating company. It licenses the toolkit from the Foundation and does the paid work. Its articles entrench the mission, the fee caps and the consent principles.
Khaos takes a fixed fee for work done and a capped share of what a model earns. The cap is 10% in the articles, and 5%, permanently, for the founding cohort. The rest goes to the people whose work made the model.
Alongside this sits the Guild of Fine Tuners, which certifies that a model's training data was the artist's to use, recorded on a public, hash-chained ledger.
Chris Green, Khaos Founder. A British composer and guitarist with twenty years in the music and games industries. Credits include the Hunt: Showdown soundtrack for Crytek with Port Sulphur Band, game audio for Electronic Arts and Criterion, and campaign music for the Premier League, BBC Sport and Jaguar. He sits on the board of PCAM and has taken part in Houses of Parliament roundtables on AI and creators' rights. He also runs Blurred Edge.
Khaos is a UK company, built in London.
The first step is usually a conversation and a listen: a ten-minute walkthrough of how the attribution and rights pieces work, available on request. Ask at khaos@khaos.dev.
The Workshop is a working prototype and the Listening Space is in proof of concept, both for public release in 2027. The toolkit is free to use.
Join the waitlist