The Institution Registry

One row in the registry is one public institution. Not one country, not one product, not one search result, and not one instance of reported AI activity. A ministry appears once whether it has adopted twenty AI tools or none; a municipality appears once whether or not anyone has ever looked at its website.

Each institution is placed at one of four levels of government. The level is not inferred from the institution’s name: it is assigned when the record enters the registry, by mapping the source’s own tier vocabulary onto a fixed standard.

National

Central government: ministries and agencies, national parliaments, and supreme or constitutional courts.

First subnational

The first tier below the centre: states, provinces, and regions.

Second subnational

The tier below that, where a country has one: counties, districts, and prefectures.

Local

The lowest tier of general-purpose and special-purpose local government: municipalities, cities, and communes, together with US school districts.

Assignment works differently for the two collection tracks, and neither one guesses. Cross-country institution types are national by construction: the compiling organisations that supply them list national bodies only. For subnational records, our research assistants record each unit against the tier vocabulary its own national register uses (regional, intermediate, or municipal), and a fixed, rule-based cleaning step maps those onto the standard: regional to first subnational, intermediate to second subnational, municipal to local. Where a source already reports a level in the standard’s own vocabulary, that value is preserved as given.

The underlying standard defines a fifth value, “other”, for units that fit nowhere else. It is deliberately reserved for rare cases and no record in the current build uses it, which is why only four levels appear anywhere on this page.

The registry is assembled from materially different kinds of source, and the difference matters more than any single count. A reader who assumes all 719,588 records arrived from one centralised list will misread the registry from the first row onward. There is no such list, and there is no single authority that publishes the world’s public institutions.

For cross-country institution types, the source is a compiling organisation: an international body that already maintains an authoritative membership roster for one kind of institution. Those rosters give the registry its national layer.

• BIS
• ITU
• WIPO
• IPU
• INTOSAI
• WCO
• IMF
• UNSD
• UNODC
• IAIS
• IOSCO
• ICN
• ICER
• IOI
• BCBS
• Egmont Group
• Global Privacy Assembly
• International IDEA

For subnational institutions there is no equivalent global roster, so these records are compiled country by country from specific national registers, official administrative lists, and statistical offices: a national local-government directory here, a census or statistical-office release there, and for some countries a cross-country data portal or boundary dataset that publishes subnational units for many countries at once. This is the larger and slower half of the work, and it is why the registry can describe local government at all.

The regenerated codebook resolves 436 distinct sources. That figure is not a count of source files, nor of the internal dataset categories, of which there are only twenty-two. It is a count of the distinct sources the coverage table actually cites, resolved by two rules: a record drawn from an organisation-level roster resolves to that organisation, so all of a body’s members count once; a record we located page by page resolves to the individual page, so each page counts once. Under those rules the 436 is 20 organisation-level rosters (the compiling organisations, plus WhoGov’s ministry roster and the NCES district file), together with 195 CIA World Factbook country pages and 221 subnational sources located country by country.

Of the 436, just under half (195) are pages of a single publication read country by country: the World Factbook’s entries for national high courts, which supply all 243 of the registry’s supreme and constitutional court records. The 221 subnational sources are the mixed set described above, not 221 national registers. The figure is an honest count of what the coverage table cites, and it should not be read as 436 independent authorities.

The registry is overwhelmingly local government by count, and that is a fact about what the registry is rather than a caveat hidden at the bottom of a page. Local-level institutions are 678,242 of 719,588 records (94.25% of everything in it). Any statistic computed over the registry as a whole is, to a first approximation, a statistic about local government. A reader interested in national ministries is working with 6,360 records, and should weight, filter, or stratify accordingly rather than treating the registry as a balanced sample of government.

Within that local tier, two distinct kinds of institution should not be run together. Municipalities (cities, towns, and communes) account for 664,901 records, or 92.40% of the registry. US school districts account for a further 13,341, or 1.85%. Together they make the 678,242 local-level total. “Local” and “municipality” are therefore not interchangeable on this page or in any analysis drawn from it: school districts are special-purpose authorities with a single policy remit, they exist in one country, and folding them into a municipality count would misstate both.

The long tail of that table is the cross-country layer: twenty-odd institution types, each with somewhere between thirty and a few hundred members worldwide, one row per country that has such a body. Those categories are small by construction rather than by omission: there are only so many central banks.

The registry is a frame of institutions, not a record of AI use. It exists to define the population that the observatory’s data collection runs against, so that coverage and absence can both be measured against something fixed. It is built and maintained entirely independently of any finding about any institution.

• Inclusion does not mean the institution has been searched.
• Inclusion does not mean anything about the institution has been reviewed.
• Inclusion does not mean the institution has been found to use generative AI.
• Inclusion does not mean the institution has been found not to use it.

The consequence runs the other way too, and it is the more important half. An institution with no activity recorded against it is not thereby an institution that does not use AI. Most often it simply reflects where collection has and has not reached. Absence of a record is not a negative finding, and it should never be read or reported as one.

A record’s evidence state (whether anything has been looked for, and what was found) is a separate matter from its presence here, and it is described in the observatory’s own terms on the parent methodology page. This page does not restate that vocabulary, to avoid a second set of words for the same thing.

Vintage

As of 2026-07-30, from master build mb-2026-07-30.

Every count on this page belongs to that build and to no other. The registry is rebuilt as new national registers are compiled and existing ones are revised, and each rebuild carries its own identifier, which is stamped onto every record it produces. A figure quoted from this page without its build is a figure with no meaning attached, because the next build will move it.

Movement between builds can be substantial and is not always a correction. The build immediately preceding this one held 706,247 institutions; this one holds 719,588. The difference is 13,341 records, and it is not an error that was fixed: it is the addition of US school districts, a category the earlier build did not cover at all. Growth in this registry usually means the frame has been extended, not that it was previously wrong.

Movement is not always growth either. On 25 August 2026 a correction moved 386 institutions out of a stronger published claim into a weaker one, with no change to the frame and nothing searched again. That correction was to the evidence states rather than to this registry, and it is described where those states are defined, under Review states on the methodology page.

Everything above is the registry in aggregate. The level of detail a reader actually needs to check our coverage of a particular country (which levels and institution types we hold for it, how many of each, and the specific register each block of records came from) runs to 2,907 rows across 225 countries and territories. That is a data view, not page copy, and pasting it here would put a table that goes stale on the next rebuild into a page that nobody rebuilds.

For this release it is published instead as the coverage codebook: a single downloadable document, regenerated from the same master build as the figures above, containing the full country-by-country tables at level × type × count, with the actual source cited inline for every block of records. It is exact, it is dated, and it stays consistent with this page because both come from one build.

Coverage codebook

Open the full country-by-country coverage tables (HTML)

Serving the same view from the observatory’s API is the better long-term answer, since it would stay current without a rebuild of this page. It is not what we are doing for this release: it would add a data-contract dependency to a page that is otherwise prose, and it queues behind other API work. It remains open as a later alternative.