Developing validity arguments for artificial intelligence-based assessment: Balancing affordances and threats.

Abstract

Background: Artificial intelligence (AI) is increasingly used to generate, score and interpret educational assessment, yet these applications are being adopted in a largely unregulated environment. This creates a paradox: Whereas AI systems used in clinical care are subject to formal scrutiny for safety, performance and monitoring, AI systems used to inform consequential decisions about learner progression and future clinical practice are not. Existing validity frameworks remain useful but may not fully account for AI-specific threats.

Methods: We conducted a state-of-the-science conceptual review examining validity for AI-based assessment through Kane's four inferences: scoring, generalisation, extrapolation and implications. We integrated conceptual and empirical literature from medical education and adjacent AI evaluation fields to identify major threats, evidence needs and practical implications for defensible AI-based assessment use.

Results: AI introduces distinct validity threats across the inferential chain. For scoring, key risks include construct contamination, prompt instability and limited explainability. For generalisation, domain shift, rater culture differences and temporal drift threaten reproducibility across settings and time. For extrapolation, AI may alter the construct being assessed, making explicit construct definition essential to defensible interpretation. For implications, risks include subgroup inequities, automation bias, deskilling, weak accountability and erosion of academic integrity. Across these inferences, we identified seven cross-cutting mechanisms that can operate as either threats or affordances depending on design and governance.

Conclusions: Validity for AI-based assessment should be approached as ongoing governance of a sociotechnical system rather than one-time validation of a static instrument or assessment process. Defensible adoption requires explicit validity arguments, auditability, multi-site evaluation and continuous consequence monitoring that is aware of the specific AI affordances and threats.