Judgment in AI Policy Advice: Humans-in-the-Loop, Appreciation, and the DOGE Experience

Recent studies of Policy Advisory Systems (PAS) focus on the quality of advice and how it is designed and managed. This article extends this analysis looking at the impact of Artificial Intelligence (AI) on advisory systems’ behaviour and activity using Geoffrey Vickers’ concept of policy judgment. It notes how negligible user costs combined with the apparent robustness of AI computational outputs can generate advice that can exercise an important and independent influence on policy decisions and outcomes, despite that advice not always being accurate or well-considered. In this sense AI-generated advice shares some of the same characteristics of earlier mechanical algorithms framing policy problems and recommendations for action which lack the hallmarks of careful appreciation of the issues at stake. This is illustrated by the use of AI tools by the US Department of Government Efficiency (DOGE) in 2025-2026, in particular its directive to autonomously reduce the number and size of contracts and grants made by many Departments, and specifically the US Department of Veterans Affairs. The case underlines that a crucial design challenge for AI-based policy advice is to go beyond only trying to ensure human oversight or a ‘Human-in-the-Loop’ in order to ensure that problems are properly articulated and actions towards them are sound.