Data licensing guide

AI training data statistics, with sources

Quick answer

The most cited figures: Epoch AI estimated about 300 trillion tokens of quality public text in 2024, likely fully used between 2026 and 2032. A 2024 audit of 14,000 domains found 5%+ of tokens in the C4 dataset, and 28%+ of its most critical sources, fully restricted from AI use. Figures on the private data market are much weaker, so we don't quote market-size estimates. Every figure here is dated and linked.

~300T
tokens of quality public text (Epoch AI, 2024)
2026–2032
likely full use of that stock (80% confidence)
28%+
of C4's critical sources restricted (2024)
45%
of C4 restricted by terms of service (2024)

Many "AI data" statistics online have no source, or quote each other in a circle. This page lists only figures we've read at source, with the date, so you can judge how old they are.

01How much public text is there?

FigureSource and date
About 300 trillion tokens of quality- and repetition-adjusted public human text (90% interval: 100T–1,000T)Epoch AI, June 2024
Fully used between 2026 and 2032, with 80% confidence, if trends continueSame paper
Around 2028 under compute-optimal training; 2027 with 5x overtrainingSame paper

A token is roughly a word fragment. These are projections, and they've moved before: the same authors' 2022 estimate was 2024.

02How much is being restricted?

FigureSource and date
14,000 web domains auditedConsent in Crisis, July 2024
5%+ of all tokens in the C4 dataset fully restricted from useSame paper
28%+ of C4's most actively maintained, critical sources fully restrictedSame paper
45% of C4 restricted by terms of serviceSame paper

The authors found the restrictions rose sharply within a single year, 2023–2024.

03What's being paid for private data?

There's no reliable market-wide figure, so we don't give one. The evidence that exists is specific:

04How we choose what goes here

A figure appears only if we've read the original, it's dated and it says what it measures. We leave out vendor market-size estimates and numbers that exist only in secondary write-ups. See how we verify figures.

05Questions people ask

Will AI run out of training data?

Epoch AI's 2024 paper projected public human text would be fully used between 2026 and 2032 if trends continued. That's a projection, and methods keep changing.

How big is the AI training data market?

We don't quote a figure. Market-size estimates vary widely and rarely show their working. The public deals tracker lists what's actually been disclosed.

Why don't you have newer numbers?

We only include figures we've read at source. If a newer primary study appears, we'll add it with its date.

Sources

Check which programmes you’re likely to matchA few short questions, about a minute. Free, no sign-up. The programme decides who qualifies.trackimpact.io/data →