The observatory

The dataset is a versioned, open-source panel of institution-level observations. It makes public-sector use of generative AI comparable across institutions and checkable against the sources it rests on.

• Long-run scope: public-sector generative AI adoption across roughly 700,000 institutions.
• Documents evidence, tools, vendors, safeguards, and governance practices.
• Publishes the data and the methods together, so that findings can be re-checked and built on.

Scope runs from national government down to local: parliaments, courts, ministries, agencies, regulatory bodies, and local authorities. Collection is phased. It starts with national institutions in every country and works down through the subnational tiers.

The full panel is still being built; the registry available today covers as far as collection has reached.

• Public statements
• Policies
• Pilot programs
• Active deployments
• Tools and platforms
• Vendors
• Safeguards
• Governance frameworks

That list is the whole of what counts, and an announcement is as much a record as a working deployment. Each documented finding carries structured metadata, the source it was drawn from, and a confidence assessment.

Fragmented evidence

Current knowledge about government AI use is scattered across news articles, press releases, and one-off reports with no consistent methodology or structure.

Geographic bias

Existing studies overwhelmingly focus on a handful of wealthy democracies. Most of the world is undocumented and under-analyzed.

Insufficient for inference

Without standardized measurement, researchers cannot make valid comparisons across countries, institution types, or time periods.

Every coded finding carries its sources: the document it came from, the passage it was drawn from, when that document was published, and what it establishes. A finding without that provenance is treated as a defect, not as a partial record. Confidence is recorded on the finding itself.

Records also carry an evidence state, so it is always visible whether an institution has been searched and what was found. Not yet reviewed means no one has looked.

Open data — still to come

Public data releases are in preparation, so that anyone can verify the dataset and reuse it.

Versioned releases — still to come

Each release will be a fixed public snapshot, numbered and with a changelog, so that a figure can be cited at a version. The release workflow is still being built.

Finding-level provenance — expanding

Institution profiles already show the sources behind each finding. Exporting that provenance in bulk is still to come.

G3O is an independent academic research project. Funding and institutional support do not influence data collection, analysis, or policy recommendations. All research outputs are subject to standard academic peer review and transparency norms.