AI System Quality as Evidence-Centric Governance: Lifecycle Accountability for High-Risk AI Systems
Conference
Regional Statistics Conference 2026
Format: IPS paper - RSC 2026
Keywords: reliability, responsibleai, transparency, trustworthiness
Session: IPS 1289 - Explainable data science
Friday 5 June 8:30 a.m. - 10:10 a.m. (Europe/Malta)
Abstract
Artificial intelligence (AI) system quality is increasingly central to responsible AI governance, particularly under Article 17 of the EU AI Act, which requires providers of high-risk AI systems to establish a Quality Management System (QMS). Standards, conformity-assessment approaches, and verification methods increasingly operationalise these requirements, but the concrete evidence needed to assess AI system quality remains dependent on the system, intended purpose, risk, and deployment context. A complementary governance problem is therefore how quality claims concerning an individual AI system remain justified as the system and its context evolve.
This paper addresses this problem by conceptualising AI system quality as evidence-centric lifecycle governance. It distinguishes AI system quality from conventional software and manufacturing quality by highlighting challenges linked to probabilistic behaviour, data dependency, context sensitivity, and temporal degradation. It then organises AI system quality concerns into four governance-oriented clusters: technical robustness and reliability, data and model integrity, human-facing intelligibility and oversight, and institutional accountability.
Building on this structure, the paper develops a framework in which quality claims are linked to context-specific and auditable evidence across the AI system lifecycle, together with responsible actors, review and challenge, monitoring, and corrective or remedial action. The contribution is a generalisable governance structure for maintaining the evidential justification of quality claims concerning individual high-risk AI systems, while allowing the concrete evidence, verification methods, metrics, and thresholds to remain system- and context-specific.