Fair GPT

The idea

A cooperative, not a scrape

The large models were built on work that was taken. The people who wrote the books, the papers, the songs and the code were never asked, and they hold nothing of what was built from them. Fair GPT is the opposite arrangement.

Members contribute their own original work to a shared corpus. The corpus is the productive asset, and the people who filled it hold it. Contribute and you get a renewing allowance of questions plus a measured share of the ledger. Contribute nothing, or contribute somebody else's work, and you are a reader who pays like any other customer.

How it works

Five things that had to be true

1. Provenance at the door. Every contribution is hashed, chunked and compared against the whole corpus for overlap. Near duplicates are held for review. You sign a timestamped attestation of authorship that can be verified later, so a false claim is attributable rather than deniable.
2. Retrieval, never weights. Contributions live in a retrieval index. Nothing is trained on them. This one decision is what makes the next three possible.
3. Withdrawal that is real. Take your work out and it leaves every future answer immediately. No model has to be retrained, and no promise has to be taken on trust.
4. Attribution per answer. Every response names the contributions it drew on, with author, title and match score. The claim is auditable on the face of it.
5. Value measured by use. Nobody grades your work when you hand it over. The ledger counts how often it was actually retrieved into an answer, and the share follows that.

Honest state

What is not built yet

It runs entirely on hardware in a flat in Munich. No cloud model is called, which keeps the running cost of an experiment about fairness honest.