Scoping Review of 4,259 Documents Finds Only 58 AI Audits Treat System Integration as a Core Evaluation Target

arXiv 2608.04921·low signalXBlueskyLinkedInCopy link

Model-centric audits miss risks that arise from interactions among components and deployment environments, a category aerospace and other safety-critical software fields have audited for decades. Scanning 4,259 documents, the authors identify just 58 AI audits that treat system integration as a core tenet, and reflexive thematic analysis of that corpus finds it fragmented: few measures target integration-specific risks, large gaps remain against traditional audit expectations, and information access heavily shapes audit design. They organize integration into three sites — inter-component, system-environment, and multi-system — each serving risk exploration, risk determination, coordination, and procedural regularity, and assessing qualities like compatibility, completeness, and oversight that component-level evaluation cannot reach.