Answers are grounded in this paper and cite their sources.
Generative artificial intelligence (GenAI) can make learning feel easier, but whether this ease helps students accurately judge what they have learned remains unclear. Drawing on the cue-utilization framework, this research examines processing fluency as a metacognitive mechanism linking GenAI-assisted learning to judgments of learning (JOLs) and calibration. Across three experiments with university students, we tested how AI assistance, explanation fluency, and unaided retrieval shape subjective learning judgments relative to subsequent independent test performance. Study 1 ( N = 264) showed that GenAI-assisted learning increased processing fluency and JOLs, with a significant indirect association linking AI assistance to JOLs through processing fluency. Study 2 ( N = 360) manipulated fluency in scripted AI explanations delivered through a simulated GenAI interface and compared them with conventional-text learning. High-fluency explanations elicited higher JOLs and greater positive calibration bias than low-fluency explanations and conventional-text learning, although the analyses did not detect statistically significant between-condition differences in objective test performance. Absolute calibration error did not differ significantly across conditions. Study 3 ( N = 386) showed that a brief no-AI retrieval attempt substantially weakened the relationship between processing fluency and JOL and reduced overall positive calibration bias. However, the retrieval-related reduction in calibration bias did not differ significantly between the high and low-fluency conditions. Together, the findings suggest that GenAI can shape metacognitive judgments by changing how fluent learning feels, while the analyses did not detect statistically significant between-condition differences in immediate independent test performance. Providing learners with opportunities for unaided retrieval may help shift judgments from AI-generated processing ease toward more diagnostic evidence of what they can independently retrieve.
No takes yet. Share an insight, caveat, or question.
Juzhi Zhang (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: