A Blueprint for Evaluating AI in Public Administration

Published Nov 9, 2026 (document metadata), from Stanford HAI's RegLab. This policy blueprint sets out how government agencies can verify the reliability of AI tools in high-stakes public services. Its central argument is that 'common AI evaluation methods are insufficient to help governments evaluate AI tools in practice' and that agencies already hold 'hidden benchmarks' — administrative records of past task performance that could be released to both measure and drive AI progress. It recommends agencies require third-party evaluations as a condition of procurement, run randomized A/B tests as systems are deployed, and collaborate with universities for cost-effective independent expertise. It distinguishes two risk categories — safety risk (tools manipulated/attacked; agentic systems acting on government data) and performance risk (failure to deliver intended benefits) — and cites the scale of deployment: 'In 2025 alone, 56 federal agencies reported 3,611 individual AI use cases within government operations, a 69% increase from the previous year.'

Government is one of the largest and fastest-growing deployers of AI, and this is a concrete procurement-and-evaluation playbook for the executive branch that also signals where third-party verification standards are likely to head — directly relevant to any vendor selling AI into the public sector.

For public-sector and enterprise AI assurance leads: test whether your procurement conditions AI contracts on independent third-party evaluation and whether you can exploit 'hidden benchmarks' in your own administrative or operational data.