Support History Shapes Metacognitive Offloading and Independent Transfer in Human–AI Learning: A Three-Arm Randomized Trial

Humans distribute cognitive control between internal effort and external resources, yet it is unclear whether metacognitive offloading induced by low-friction artificial-intelligence support is reversible. In a three-arm, individually randomized multisite trial, 486 undergraduates completed 32 reasoning tasks across 16 sessions. Access to complete solutions was switched after Session 8 from solution-available to retrieval-contingent (SA→RC), held retrieval-contingent throughout (RC→RC), or held solution-available throughout (SA→SA). Independent transfer was assessed without model access at four occasions across 20 weeks; session-level records yielded an objective metacognitive offloading index. At the delayed assessment, SA→RC outperformed SA→SA by 5.14 points on a 0–48 criterion, 95% CI [2.88, 7.40], g = 0.42, and did not differ from RC→RC. Under identical Phase 2 support, however, SA→RC retained greater offloading than RC→RC, Δ = 0.40 SD, 95% CI [0.28, 0.52], g = 0.43. Verification and correction after error recovered to the RC→RC level, whereas confidence–performance calibration neither differed by arm nor preceded behavioral reorganization. Interviews with 36 participants converged on unmonitored availability in SA→SA and a reappraised enforced pause in SA→RC. Thus, performance recovery and control-allocation recovery were separable: redesigning access restored later unassisted performance relative to continued solution availability, while prior support history remained behaviorally detectable.