RADAR is out. I am one of its three authors, so every disclosure applies. With Luke Jordan and Manuel Ramos-Maqueda, we tested two things across 166 countries: whether an AI system can explain how to get a public service, and whether it can then reach the official entry point to begin.
It does the first far better than the second. Conventional rankings, the UN’s E-Government Development Index among them, measure the explaining reasonably well and largely miss the reaching. A government can sit high in those tables and still present an agent with unstable links, JavaScript-heavy pages and bot controls nobody chose on purpose.
Most of the repair is dull: stable URLs, content a crawler can read, and navigation that survives a redirect.
The layer between government systems and the AI tools citizens already use looks like digital public infrastructure. Identity, payments and data exchange were built because they were shared, dull and underneath everything else, and what the paper calls sovereign legibility is dull in the same way. It is also the first piece of the stack a government forgets, because the agents arriving at the door do not announce themselves and mainstream analytics does not show the traffic they generate.
Who pays for plumbing that is more useful and duller than the deployments sitting on top of it?
Assorted links
- A new paper on concurrent agent systems names the failure mode procurement will meet first: authorisation granted when a request is made is already stale by the time the effect executes.
- Stanford’s AI Index counts AI witnesses at US congressional hearings rising from 5 in 2017 to 102 in 2025, a twentyfold increase in who gets asked.
- From September, Brazil’s federal IT contractors may not decide on their own to use AI on government systems, a rule aimed at how the tools arrive rather than at what they produce.