Commitment, Not Sequence: A Formal Model of Error Foreclosure in AI-Mediated Learning

Abstract

We propose and formalize a specific cognitive mechanism — error foreclosure — for how generative AI reshapes learning, and distinguish it sharply from cognitive offloading. Offloading concerns processes relocated to an external store (memory, calculation); error foreclosure concerns a process that never begins. Our claim is that durable model construction depends on self-generated prediction error, which requires a falsifiable epistemic commitment — a determinate, staked position that can be recognized as wrong when external content arrives. Answer-first AI use can remove that commitment before it forms, and therefore forecloses the error that drives learning. This single mechanism, and nothing broader about AI-assisted learning, is what the paper defends. A growing body of work shows that people learn less durably when generative AI supplies answers early in a task than when they attempt it first — visible in productive-failure research, pretesting, cognitive forcing functions, and AI-tutoring field experiments. The dominant reading treats this as a timing or engagement effect. We argue that neither timing nor engagement is the causal ingredient; commitment is. We formalize this as a Bayesian rule-learning model in which commitment modulates processing depth through registered surprise, and contrast it with an activation model in which engagement alone suffices. Across parameter sweeps the two mechanisms produce opposite signatures on a four-condition design that dissociates commitment from engagement, and their mixture is identifiable from the resulting contrasts. The model further yields predictions absent from verbal theory: an expertise attenuation-without-reversal constraint, a transient underconfidence after committed learning, a hysteresis prediction that friction interventions will show a prevention-versus-cure asymmetry, a competence-gated protection effect whereby committing first shields learners from absorbing AI errors only in proportion to their existing competence, and a diagnostic prediction whereby the sign of the misconception effect reveals whether learners update by surprise-driven correction or by belief perseverance. We specify the human experiments that would confirm or refute each claim, and pre-state the conditions under which the mechanism should be rejected in favor of simpler existing accounts. STATUS AND SCOPE: This is a theory-and-computational-modeling preprint. It reports simulation results only and makes no empirical claim about human participants. Section 8 pre-states six results that would falsify the account. The core hypothesis is the author's own; the commitment-versus-engagement dissociation emerged jointly in dialogue with AI systems; the formalization, simulation code, and manuscript drafting were produced with AI assistance, and the author is not able to independently verify every step of the formal derivation. The formal results should therefore be treated as requiring independent verification, and the complete code is released with a single-command reproduction script (run_all.sh) so that verifying them is cheap. A pre-submission audit of the author's own code overturned one result that an earlier version had flagged as non-trivial; this is documented in the changelog. All claims are the author's responsibility.