Governing Risks Arising from Institutional AI-Driven Decision Support Practices: A Case of Women Leaders in Tertiary Education

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The increasing integration of large language models (LLMs) and generative artificial intelligence (GenAI) into tertiary education is extending artificial intelligence (AI) from teaching and learning toward institutional decision support, creating new requirements for human oversight and risk governance. This study examines an integrated behavioural risk-governance mechanism comprising Trust in LLMs (TR), AI Overdependence (AIO), AI-Driven Decision Risk (AIDR), Verification Behaviour (VB), and Responsible AI Decision Governance (RAIDG). Using a cross-sectional survey, data were collected from 849 women leaders representing 12 nationalities in tertiary education organisations in Saudi Arabia who had prior or current experience with LLM- or GenAI-supported institutional decision-making. The data were analysed using Partial Least Squares Structural Equation Modelling (PLS-SEM). TR was positively associated with AIO (β = 0.534, p < 0.001), while AIO was positively associated with AIDR (β = 0.408, p < 0.001) and negatively associated with VB (β = −0.521, p < 0.001). AIDR was positively associated with VB (β = 0.208, p < 0.001), whereas VB showed the strongest positive association with RAIDG (β = 0.620, p < 0.001). AIO (β = −0.198, p < 0.001) and AIDR (β = −0.143, p < 0.001) were negatively associated with RAIDG. Indirect association analysis further showed that TR was positively associated with AIDR through AIO (β = 0.218) and negatively associated with VB through AIO (β = −0.279). The serial indirect pathway from TR to RAIDG through AIO and VB was also significant and negative (β = −0.173), supporting the hypothesised governance-risk mechanism. The model explained 52.0% of the variance in RAIDG and demonstrated positive out-of-sample predictive relevance, with RAIDG showing the highest Qpredict2 value (0.512). Based on these findings, the study proposes a Behavioural Governance Operating Model (BGOM) integrating trust calibration, overdependence monitoring, decision-risk assessment, risk-adjusted verification, human decision authority, accountability, and continuous monitoring. The study extends AI-in-education research from adoption toward post-adoption behavioural risk governance. It identifies verification as a central human-in-the-loop mechanism for responsible institutional AI decision-making.