Testing the applicability of a governance checklist for high-risk AI-based learning outcome assessment in Italian universities under the EU AI act annex III

Background The EU AI Act classifies AI-based learning outcome assessment as high-risk (Annex III, point 3b), yet sector-specific frameworks for institutional self-assessment remain underdeveloped. This creates accountability gaps affecting student rights and educational equity, as institutions lack systematic tools to demonstrate that algorithmic assessment systems produce valid and fair outcomes. Methods This exploratory study tests whether ALTAI’s trustworthy AI requirements can be operationalized for educational assessment governance through the XAI-ED Consequential Assessment Framework, which integrates three educational evaluation theories (Messick’s consequential validity, Kirkpatrick’s four-level model, Stufflebeam’s CIPP). Following pilot testing with three institutions, four independent coders applied a 27-item checklist to policy documents from 14 Italian universities (13% with formal AI policies plus one baseline case) using four-point ordinal scoring and structured consensus procedures. Results Intercoder reliability analysis revealed substantial agreement (Fleiss’s κ = 0.626, Krippendorff’s α = 0.838), with higher alpha reflecting predominantly adjacent-level disagreements suitable for exploratory validation. Analysis of 14 universities reveals substantial governance heterogeneity among early adopters (Institutional Index: 0.00–60.32), with Technical Robustness and Safety showing lowest implementation (M = 19.64, SD = 21.08) and Societal Well-being highest coverage (M = 52.38, SD = 29.38). Documentation prioritizes aspirational statements over operational mechanisms, with only 13% of Italian institutions having adopted AI policies by September 2025. Discussion The framework demonstrates feasibility for self-assessment but reveals critical misalignment: universities document aspirational commitments more readily than technical safeguards, with particularly weak capacity for validity testing and fairness monitoring. Findings suggest three interventions: (1) ministerial operational guidance translating EU AI Act requirements into educational contexts, (2) inter-institutional capacity-building addressing technical-pedagogical gaps, and (3) integration of AI governance indicators into national quality assurance systems to enable systematic accountability. The study contributes to understanding how educational evaluation theory can inform the translation of abstract trustworthy AI principles into outcome-focused institutional practices under high-risk classifications.