Methodology

Per-dataset caveats live with the dataset, in the popover attached to every figure and on the sources page. Terms are defined in the glossary. This page covers the choices that apply across all of them.

Our own indexes

• Governance Readiness Index: the scoring rubric, what it does and does not measure, and how the EU is handled.
• AI Law and Policy Index: how instruments are selected and how "how binding is this" is decided.
• AI Usage Disclosure Index: token volumes, user counts and shares of use as each company or study states them, with the scope of each. Rates are converted to tokens per month here; an annual or quarterly total becomes a monthly average and is marked as one.
• Environmental Disclosure Index: what each developer and data-centre operator has published about training energy, emissions, sites, power sources and energy per prompt. Each entry is marked as published by the developer, partly published, reported by others, or not disclosed.
• AI Datacentre Index: AI datacentre sites that Epoch AI's hub does not hold, from the statements of developers, named tenants and planning authorities, starting with Australia. A site qualifies only where a primary source says it is built for AI compute or names the AI deployment in it. Capacity is as stated, never estimated or converted between IT and facility power. A site within two km of one Epoch tracks is treated as Epoch's.
• Release Safety Disclosure Index: every release Epoch AI lists from the developers in the Frontier Safety Framework Index since January 2025, against the system card, model card or safety report the developer published for it. A card is on the day if dated no later than the day after Epoch's publication date. Results means the card reports chemical, biological, cyber or autonomy evaluation results for that model; a card relying on another model's results is marked as such. Government testing counts only an institute the card names as having tested a pre-release version of that model. A threshold is reached where the card says so under the developer's own framework, or says it could not be ruled out and applies the safeguards it requires. Dated updates to earlier models, embedding models and single-task models are listed and not counted.

Comparing things that are not comparable

Several pages set two measures against each other. Where the measures have different units, they are converted to percentile ranks within the set of entities present in both before being compared. A use rate of 40 percent and a readiness score of 40 are not the same kind of 40, and subtracting one from the other directly would produce a number with no meaning.

The consequence, stated wherever it applies: the picture moves when the set of entities changes. A rank is a statement about relative standing, which is all a cross-unit comparison can honestly be.

Trends

Where a growth rate is fitted, it is ordinary least squares on the logarithm of the quantity against time, over a window that is stated. It describes what has already happened and is deliberately not extended past the last observation. This site publishes no forecasts of its own.

Fits are checked against alternative windows and the result of that check is printed in the figure's own note, so a reader can see whether the answer depends on where the window starts.

Missing data

A suppressed or absent value is never treated as zero. An entity with no data is left unshaded on a map and absent from a table, and the count of what is missing is stated next to the count of what is present. Where a join drops entities, they are named.

This matters more than it sounds. Reported compute clusters, reported incidents and reported adoption all measure reporting as much as they measure the thing. An absence of incidents in a country usually means an absence of reporting.

Freshness

Every figure carries the date its data was retrieved, and anything older than 45 days is marked stale rather than presented as current. When a source cannot be fetched, the last cached copy is served and its age is shown: stale data with a visible date beats a page that vanishes.

A dataset is only rewritten when its records change, so the repository history distinguishes a day when data moved from a day when the pipeline merely ran.

Classification

News items are categorised by keyword rule, and every item on the news pages prints the terms that matched. The rule is blunt: it misses things and occasionally files something oddly. Publishing the terms is what makes it arguable rather than authoritative.

Checks

Each stage carries a runnable check, and the ones that matter are verified by being deliberately broken to confirm they fail. The licence guard is checked against sources it must block. The news classifier is checked against real headlines it previously got wrong. The map's coordinate decoder is checked against a hand-computed example and against the mistake that would otherwise be made silently.

That last habit came from a real incident: a bulk edit once rewrote a normalisation rule and the assertion guarding it into a matching pair that passed while doing nothing. A check that still passes when the logic is broken is worse than no check.

Typography

Extracts from feeds have their dash characters normalised to plain hyphens, which is a house style rule rather than an editorial one. It changes presentation and not meaning, and the unaltered original is one link away from every extract.