Human-centric AI governance in the European Union. Accountability, fundamental rights, and institutional resilience in public administration
TL;DR
This Policy and Practice Review concludes that AI-enabled administration remains legitimate only where public authorities retain the capacity to understand, justify, correct, suspend and democratically control the systems they use.
Abstract
Artificial intelligence is increasingly used in public administration to classify individuals, assess risks, prioritize cases, support eligibility determinations and guide the allocation of public resources. In the European Union, these uses are governed by the Artificial Intelligence Act, the GDPR and the Charter of Fundamental Rights. Formal compliance, however, does not by itself ensure lawful and accountable administration. AI relocates discretion from the visible act of decision-making to less visible choices concerning data, model design, procurement, thresholds and interface architecture. This Policy and Practice Review therefore treats human-centric AI governance not as a general ethical aspiration, but as an administrative and constitutional framework for governing public power. Drawing on EU law, public administration scholarship and a comparative institutional analysis of selected Member State practices, it develops six interdependent dimensions: legal anchoring, accountable discretion, fundamental rights by design, meaningful human oversight, contestability and justification, and institutional resilience. The analysis shows that common EU rules may produce unequal levels of protection where public authorities differ in technical expertise, audit capacity, procurement independence and access to effective remedies. It also argues that accountability must follow the chain of influence through which algorithmic systems shape administrative outcomes, rather than only the formal chain of decision-making. The article translates this framework into actor-specific recommendations concerning fundamental rights impact assessments, procurement, auditability, human oversight, transparency, contestability and post-deployment monitoring. It concludes that AI-enabled administration remains legitimate only where public authorities retain the capacity to understand, justify, correct, suspend and democratically control the systems they use.