The Global Index on Responsible AI marks 135 countries against 38 indicators. Transparency and explainability is one of its strongest indicators, with 58% of countries having some framework. Only 18% require public disclosure of the algorithms the government itself runs.
Of the 135, 126 have at least one government policy or initiative across the index’s 17 AI-policy indicators. Nine do not. That count has stopped telling anybody apart, and it will keep climbing.
The 18% counts requirements, not sight. A government may already publish what it runs under freedom-of-information rules, with no AI-specific mandate. The figure marks where a duty has been written down.
There is a harder ceiling, upstream of any rule. A register can only hold what the state bought and admits owning. A UNESCO survey of more than 560 judicial workers across 96 countries found 44 percent of them using AI tools at work, and 9 percent whose organisation had issued guidelines or training. Those tools arrived through the individual, not through procurement, and leave no contract to enter in a register.
So the part a register can describe is shrinking against the rest: 44 percent of those judicial workers are already using tools nobody procured. And a register names systems, not why a claim was refused, or what the person can do next.
Assorted links
-
Sarah Gold on Britain scrapping its Digital ID programme: the objection was to a scheme announced without warning and presented as mandatory, not to the identity layer, which is what an AI agent needs to act for anyone.
-
Campaigns in the United States are trialling focus groups of AI-generated voters, and a researcher there notes models emulate average behaviour well and outliers badly, which is who a public service is for.
-
Google’s AI Overviews cut daily traffic to English Wikipedia by roughly 15%, measured against the Hindi, Indonesian, Japanese and Portuguese editions, so a ministry counting page views is watching a channel close.
-
Two internal-control units in Brazil’s Federal District report processing times falling 18.2% and 50% after staff training on free models, and the paper claims only that the gains accompanied it.