---
title: "Data Credits > Carbon Credits"
description: "Carbon credits failed at the one thing data credits are naturally good at. A file has a hash, a training run is a logged event, and a public ledger turns \"nobody is checking\" into a receipt anyone can read."
canonical: https://www.askgib.com/articles/data-credits
last-updated: 2026-09-07
lang: en
---
# Data Credits > Carbon Credits
> Carbon credits failed at the one thing data credits are naturally good at. A file has a hash, a training run is a logged event, and a public ledger turns "nobody is checking" into a receipt anyone can read.
By Gibril Bachouchi. Filed 2026-09-07. 4 min read. Web page: https://www.askgib.com/articles/data-credits
Languages: [en](https://www.askgib.com/articles/data-credits.md) · [fr](https://www.askgib.com/fr/articles/data-credits.md) · [ar](https://www.askgib.com/ar/articles/data-credits.md) · [es](https://www.askgib.com/es/articles/data-credits.md) · [ru](https://www.askgib.com/ru/articles/data-credits.md) · [pt](https://www.askgib.com/pt/articles/data-credits.md) · [it](https://www.askgib.com/it/articles/data-credits.md) · [de](https://www.askgib.com/de/articles/data-credits.md) · [ja](https://www.askgib.com/ja/articles/data-credits.md) · [ko](https://www.askgib.com/ko/articles/data-credits.md) · [hi](https://www.askgib.com/hi/articles/data-credits.md) · [he](https://www.askgib.com/he/articles/data-credits.md) · [zh](https://www.askgib.com/zh/articles/data-credits.md) · [tr](https://www.askgib.com/tr/articles/data-credits.md) · [kab](https://www.askgib.com/kab/articles/data-credits.md)
They tell you the service is free.
In 2021 Apple added one switch, asking people whether they wanted to be tracked. Meta told investors it would cost them $10 billion the following year. One data source, one platform, one company, one year.
That is the price of the thing you were handed for free. It is also the only honest number anyone has ever published about it, and it exists only because a competitor forced it out.
Meta has paid $5 billion to the FTC over Cambridge Analytica and €1.2 billion for the largest GDPR fine ever issued. Its business is bigger now than before either.
A penalty you can budget for is not a penalty. It is a price. Twenty years of regulation has produced a cost of doing business and a consent banner, and the money still runs one way. The person who generated the data gets a free service. The company that collected it gets the hundreds of billions.
The problem is not that the rules are too weak. It is that nobody verifies anything.
You make an account. You accept something. What happens to what you do next, who it is sold to, which model it trains, which broker it ends up in, is disclosed nowhere you will ever look, and nobody audits whether the policy you accepted matches what the company actually did.
It is not that the answer is bad. There is no answer, because nobody is asking the question.
Twenty years of it is already inside the models. It cannot be taken out. Unlearning does not work and no court is going to order a frontier model deleted, so every year spent litigating backwards is another year the same extraction runs forwards, unpriced.
Stop trying to claw it back. Price it going forward. That is what a credit is for.
Carbon had the right idea and the wrong substrate.
**It got the unit right.** A tonne of CO2 equivalent is one number, measurable, the same in Lagos and in Lyon. You can write it into a contract, a tax code, an audit and a court filing. That unit is the only reason a carbon market could exist at all.
**It got the audit wrong, and the audit is what killed it.** A forest is in a place. To check it you have to travel there. What you are measuring is a counterfactual, meaning what would have happened otherwise, and nobody could afford to find out. Junk offsets flooded in, the certificate stopped meaning anything, and the price signal died.
Now run those same two tests on data.
**The unit is easier, not harder.** The standard objection is that data has no unit, and that is the most dangerous idea in this debate. No unit means no rule, no rule means nothing to enforce, and the whole thing dies as a white paper. Complexity is not neutrality. It is a decision to do nothing, made slowly. So pick one and publish it: the registered work, weighted by category. A novel is not a tweet. Coarse and workable beats elegant and theoretical.
**The audit is trivial, not impossible.** A registered work is a file with a hash. Its use is a database event inside a training pipeline that already logs everything for its own engineers. Nobody has to travel anywhere. Nobody has to prove a counterfactual.
So put the credits on a public, append-only ledger. Registering a work mints one. A training run above the threshold burns one. The receipt is permanent and readable by anyone, without asking permission from the lab or the regulator. Train on unregistered data, or buy a dataset from someone who did, and you owe what you dodged times three or five, whatever makes the honest route cheaper.
That is the pitch. Carbon credits failed at exactly the thing data credits are naturally good at. And a public burn record is the answer to "nobody is checking," because it is a disclosure that nobody has to be trusted to make.
Set the levy the way a radio station pays a blanket licence to broadcast: a percentage, above a compute threshold, into a fund.
| Levy rate | On $100B of model revenue | On $200B | On $500B |
|---|---|---|---|
| 1% | $1.0B | $2.0B | $5.0B |
| 3% | $3.0B | $6.0B | $15.0B |
| 5% | $5.0B | $10.0B | $25.0B |
Call it **$6 billion a year**, and it grows with the industry automatically, which is the point of a percentage.
This is the trap every "pay the creators" proposal walks into.
Weight the works, a novel counting 1,000 units against a forum post's 1, and a plausible registered pool comes to roughly 60 billion units. Six billion dollars across 60 billion units is ten cents a unit.
A novel earns **$100 a year**. A photograph earns **$1**. A forum post earns **ten cents**. It gets worse annually, because registration only grows. You would spend more administering a $1 payment than the payment is worth, which is why streaming is an insult to working musicians.
So do not pay per head. Let attribution decide where the money goes, then spend it as a fund. Half to the sectors, languages and countries the training actually drew from. The rest paying people to release good work under a training licence, commissioning what the market underfunds, and covering the audits, so enforcement scales with the industry instead of waiting for a ministry to find budget.
For scale: $6 billion a year is 60,000 grants of $100,000. The entire public arts funding of the United States, the United Kingdom and France put together is a fraction of it.
Attribution is crude. One country's levy is another country's export tax. A collecting body is a monopoly and needs watching.
None of that is why it has not happened. It has not happened because the people making hundreds of billions prefer the current arrangement, where the service is free, the fine is a line item, and nobody is checking.