I think we're heading toward a future where anyone training an AI model will need to show where their training data came from — who owned it, what the terms were, what was paid, and when. Not as a courtesy. As a requirement: from regulators, from courts, and from customers who don't want to buy a model with a lawsuit inside it.

Software already went through this. Twenty years ago, nobody tracked which open-source components were inside a product. Then came license lawsuits, then security scares, and now a "software bill of materials" — a simple list of every ingredient in a piece of software — is standard in procurement and written into regulation. Training data is on the same path, a few years behind. The end state is a bill of materials for models: a verifiable record of everything that went in.

For that record to mean anything, you need a chain with four links. Each one exists today in some form, but almost nobody has connected them.

Layer 1: Saying the terms

The chain starts with the content owner stating their terms in a way a machine can read.

For decades the only tool was robots.txt, which can only say "crawl" or "don't crawl." That's not a license — it's a door sign. Really Simple Licensing (RSL) is a newer attempt to do better: an XML file that says what uses are allowed, what they cost, and how to pay.

This site publishes one at /license.xml. It says: search and indexing are free with attribution; AI retrieval (an agent reading my writing to answer someone's question) is metered and paid; AI training rights aren't sold automatically at all — those require an actual agreement.

The intellectual ancestor here is Datasheets for Datasets, a 2018 paper that proposed something radical at the time: every dataset should come with documentation about where it came from and how it may be used. RSL is essentially that idea, made machine-readable and priced.

Layer 2: Paying, and getting a receipt

Terms alone are just a sign on the wall. The second link is evidence that a specific party actually acquired specific content under those terms at a specific time. This is where money leaves a mark — and it's the emptiest layer in the whole chain today.

Here's how it works on this site. When an AI agent requests my content through the API without paying, it gets HTTP 402 — "Payment Required" — along with everything it needs: an address, a price, and the license terms. It pays, submits the transaction ID, and gets back a token good for a metered number of requests, plus a statement of exactly what the payment licensed: retrieval use only, attribution required, no redistribution, training not included.

The blockchain part isn't ideology. It's that a payment on a public ledger is a timestamped record that anyone can verify independently, without trusting me or the payer. The transaction ID plus the scope statement is a primitive receipt — the atom a future bill of materials would be built from.

Layer 3: Keeping the paper trail attached

The third link is where the chain usually breaks.

Content rarely goes straight from a website into a model. It gets scraped into datasets, merged into bigger datasets, filtered, deduplicated, and repackaged — and at every step, the license information tends to fall off. The Data Provenance Initiative audited more than 1,800 popular training datasets and found license information missing over 70% of the time, and wrong more than half the time when present. The terms existed at the source. They just didn't survive the trip.

The likely fix is provenance that travels inside the content rather than alongside it. The best-developed model is C2PA, a standard from the media world that embeds signed, tamper-evident records in a file — who made it, what was done to it — that survive as the file moves around. Something like that, applied per-document to licensed text, is what would let a receipt from layer 2 still be attached when the content lands in a training run.

Layer 4: Proving what actually happened

The last link is verification: can anyone confirm what a model was actually trained on?

Research is attacking this from both ends. The cooperative end: Tools for Verifying Neural Models' Training Data proposes "proof of training data" — protocols that let a model trainer prove to an auditor which data produced a set of model weights. The adversarial end: Copyright Traps for Large Language Models shows that a content owner can plant unique, invisible sentences in their text and later detect whether a model trained on them — a canary in the data mine.

Honest terms at the source, receipts at the transaction, manifests that survive transfer, and proof at the end. That's the full chain.

Why now

Three forcing functions are converging. The EU AI Act requires providers of general-purpose models to publish summaries of their training content. Copyright litigation is making "what did you train on?" a discovery question with legal consequences. And enterprise buyers are starting to ask for models with clean data pedigrees the way they already demand clean software bills of materials.

Meanwhile the free-for-all is ending on its own: Consent in Crisis documents how quickly websites are locking down against AI crawlers. When unprovenanced data becomes a liability, data with receipts becomes the premium product.

What this site is demonstrating

To be clear about intent: this is a demonstration, not a business. My blog's words are not going to fund my retirement at 10,000 satoshis per grant.

But most of the work on this problem lives at layer 1 (declaring terms) or layer 4 (catching violations), while the middle of the chain — actual transactions with actual receipts that could survive into an audit — is nearly empty. A personal website is a fine place to show the middle works: the whole loop, from machine-readable terms to paid access to a scoped receipt, running end to end at the smallest possible scale.

If it works for one blog, the mechanics work. Scale is someone else's demonstration.

Reading list

• Datasheets for Datasets — Gebru et al., 2018. Where dataset documentation norms started.
• The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI — Longpre, Mahari et al., 2023. The evidence that license information doesn't survive aggregation.
• Tools for Verifying Neural Models' Training Data — Choi, Shavit, Duvenaud, NeurIPS 2023. Proof-of-training-data protocols.
• Copyright Traps for Large Language Models — Meeus et al., ICML 2024. Detecting unlicensed training from the content owner's side.
• C2PA Specifications — the most mature standard for provenance that travels embedded in content.
• Consent in Crisis: The Rapid Decline of the AI Data Commons — Longpre et al., 2024. How fast the open web is closing to crawlers.

---

Feedback welcome — especially from anyone building the layers I didn't. The machine-readable terms are at /license.xml, and the payment flow is live on the API.