When Is Checking Enough? Verification Sufficiency Calibration as a Missing Construct in Generative-AI-Supported Higher Education

Abstract

Generative artificial intelligence (GenAI) has made verification a central educational concern. Current research increasingly asks whether students check AI-generated claims, how well they verify them, whether their trust is calibrated, and whether verification supports learning. A less developed question concerns the termination of verification: how does a learner decide that the available evidence is sufficient and that checking can reasonably stop? This conceptual article introduces Verification Sufficiency Calibration (VSC), defined as the degree to which a learner's decision to stop verifying a GenAI-supported claim is proportionate to the evidential demands, uncertainty, consequences, and disciplinary standards of the task. A targeted integrative review of recent research on GenAI literacy, verification behaviour, epistemic vigilance, evaluative judgement, trust calibration, and epistemic dependence was conducted through 3 October 2026. Recent work has documented task-sensitive verification thresholds and explicit 'enough checking' rules; it has distinguished verification initiation, quality, success, and reliance calibration; and it has developed broader AI-literacy and assessment-literacy measures. However, the adequacy of the stopping decision itself remains under-specified as a distinct educational construct. We propose a VSC framework that distinguishes under-verification, calibrated verification, and over-verification using task-specific acceptable evidence bands rather than a universal number of checks. Six propositions specify expected relations among VSC, task stakes, accountability, time pressure, prior knowledge, source cues, and transfer. We further propose a performance-based assessment architecture combining instrumented tasks, evidence-state scoring, stopping rationales, reliance decisions, and delayed AI-withdrawal transfer. VSC extends, rather than replaces, GenAI literacy, epistemic vigilance, evaluative judgement, and trust calibration by identifying an educationally consequential regulatory decision that lies between verification activity and subsequent reliance. The framework shifts responsible-AI education from the maxim 'always verify AI' towards a more demanding competence: knowing what requires verification, how much evidence is enough, and when continued checking no longer improves epistemic responsibility.