Methodology

Data collection pipeline

Structured, AI-augmented web searches using institution-specific queries; LLM-assisted extraction of structured fields from unstructured web content.

Evidence standards

Each coded finding requires at least one verifiable source document. Each source carries its own credibility assessment and each finding a confidence level, with a distinction maintained between stated intent and verified deployment.

Classification logic

Use type, deployment status, and governance measures are coded against a standardized taxonomy. Institution coding is hierarchical and aligned with international standards. Classification is machine-assisted, with human validation for ambiguous cases.

Coverage codebook

Full country-by-country coverage tables for the institution registry (HTML)

Pipeline report

Step-by-step method and measured operational benchmarks for the collection pipeline (HTML)

The unit of record is an institution, not a country and not a product. Each record describes one public body and what is known about its use of generative AI: whether it uses it, under what policy, with what safeguards, and through which procurement route.

Provenance attaches to a coded finding, not to a record. Every finding is published with the sources behind it, shown next to the finding on the institution profile rather than held in an internal file. Registry facts such as country or institution type carry no source, because the project makes no coded claim about them.

• The source document, linked.
• The passage the finding was drawn from, quoted.
• What the source establishes: whether it confirms the activity, confirms its absence, is ambiguous, or is background only.
• The date the source was published.
• The source’s type, language and credibility.

Confidence is recorded on the finding rather than on each source, because one finding can rest on several sources of differing quality.

Records carry one of four evidence states: documented activity, possible AI use, no documented activity, or structure only and not yet reviewed. Not yet reviewed is the most common. It means the institution is in the frame because the master list put it there and no search has been run against it. It is not a claim that the institution does not use generative AI. That claim is “no documented activity”, and it is only made once a search has reached the institution, whether or not that search recorded a result.

Possible AI use means the search did not settle whether the institution uses generative AI. Usually it found pages that might point to generative-AI use and none that establish it, and those pages are listed on the record. It also covers a search that stopped before reaching a conclusion but still recorded a result, which may have no pages to list. It is not a finding.

An institution whose search reached a conclusion with no page recording generative-AI use is published as no documented activity, even when one of the pages could not be read. So is one reached by a run that recorded no result at all.

The second figure on the homepage counts documented activity and possible AI use together. Possible AI use is not a finding, so that figure says how many institutions have documented or possible generative-AI use, not how many use it.

The third figure on the homepage, documented use cases, counts every finding recorded across all institutions. An institution with several findings is counted once for each, and findings are not deduplicated.

Until 25 August 2026 these states were published differently, and 386 institutions were affected. Each had been published as no documented activity, which on this site asserts that a search was run and found nothing. For those 386 the search had not established that. Some had been assessed as unclear by the pipeline and the assessment was lost when the results were loaded; for the rest the run did not finish. From 25 August they were published as unclear. Of those, any the pipeline had judged “no” on every page it could read, though the run did not finish, are now published as no documented activity, as described above. The error was found by hand, during checks made before a later load, and it had been present since the first one.

Nothing was searched again, and no finding was added, removed or revised. The number of institutions with documented activity did not change, and neither did the number screened: the correction moved 386 records from a stronger claim to a weaker one, and described the same evidence correctly. The About page commits that absence of evidence is shown as absence and never as a negative finding, and for these records the site had not been keeping that commitment.

Observations are versioned rather than overwritten, so a record can be cited at a point in time and later corrections remain visible.

What is collected

We collect only publicly available information about government institutions and their technology use. We do not scrape private or classified information.

Privacy considerations

Government personnel data is handled carefully. We document institutional roles, not personal information about individual civil servants, unless they hold public-facing leadership positions.

Uncertainty and limitations

We document uncertainty in our data: confidence scores, evidence quality flags, and limitation disclosures are built into every record.

Documentation norms

Methods, models, and decisions are logged and published alongside the data, so the project documents its own process as well as its subject.