Making public sector AI accountable

Registers were built independently, with different fields, categories, and ideas of what counts as a disclosure. So before building anything, we compared them.

We went through each register field by field, asking which fields matter for accountability work — vendor, human oversight, appeals, impact assessment, deployment scale — and which registers actually provide them. That comparison produced a framework: a common structure for what a record about a government AI system should hold, and a controlled vocabulary for each field. We then mapped each register's own terminology onto that framework — which of their fields corresponds to which of ours, and how their values translate into our vocabularies. That mapping is what the extraction follows.

The framework and the crosswalk tables — lifecycle status, domain, algorithm type, risk classification — will be published in full. We'll add them here when they're ready.

We collect records from each registry automatically. An LLM then extracts the content into our framework, following the terminology mapping described above — it's applying a mapping we built by hand. Everything is labelled: source text and AI-generated content are visually distinguished throughout, so you can always tell which you're reading.

And everything links back: every record carries a link to its source, so you can check it against the original. We review output through random sampling.

Registration is voluntary in almost every country here, so what's on this platform is largely what governments choose to disclose. A system's absence from PublicAlgo is not evidence that it doesn't exist.

Coverage is uneven. Some countries in our scope have no public registry at all and so appear nowhere in the platform yet.

Some content is machine-extracted and may contain errors. An LLM maps each register's fields onto our framework. We review output through sampling and cross-register comparison, but not every record has been checked by a person. AI-generated content is labelled throughout, and every record links to its source. See where we use AI.

Mapping between registers is a judgement, not a translation. Fields that look equivalent often aren't — a Dutch lifecycle status and a UK phase field are close but not identical. Where the mapping is approximate we flag it, but comparing across countries means accepting some imprecision. See how we compare registers.