Public Service AI Governance Frameworks Risk Failing with General-Purpose AI

A new paper argues that existing public service AI governance frameworks are ill-equipped to handle general-purpose AI (GPAI) built on large language models. The authors contend that the very properties making GPAI attractive—generality, accessibility, low deployment cost—undermine the conditions for AI safety. Traditional safety concepts like accuracy, bias, explainability, and accountability were designed for narrow, purpose-built AI and cannot be applied to GPAI's unbounded outputs. For example, accuracy cannot be quantified over unbounded outputs, and bias cannot be disaggregated when outputs are free-text judgments. The paper draws lessons from policing to illustrate these failures. Published on arXiv, the study highlights a critical gap in governance as public services increasingly adopt AI to meet rising demand with falling resources.