Abstract
Public-sector AI is often discussed as a matter of systems, models, or automated decisions. This chapter shifts the focus to the administrative outputs that people encounter: debt estimates, suspicion profiles, scrutiny pathways, routing categories, and risk reports. It argues that governance problems arise when these outputs enter public workflows and begin to operate as state-backed claims or burdens before legality, reasons, contestability, and accountability are secured. Through Robodebt, the Dutch childcare benefits scandal, UK visa streaming, and SyRI, the chapter shows how uncertain or opaque outputs can become public harm before any final automated decision is made. It develops an impact-centered framework across individual, procedural/legal, social/distributive, and institutional dimensions, and proposes an impact-governance framework based on legality before automation, traceable reasons, contestability, human accountability, and post-deployment audit. The chapter contributes an output-centered approach to governing AI creation systems in administrative practice.

This publication has 23 references indexed in Scilit: