Verification without criteria

Generative AI performs the analytical and evaluative tasks previously managed by learners, and the field has responded by designing scaffolds that restore monitoring. However, whether monitoring represents the primary site of regulatory failure remains untested. To address this, we conducted a systematic review and Epistemic Network Analysis (ENA) of 65 empirical studies (2022–2026). Using the published study as the unit of analysis, we examined the relational structures connecting instructional designs and regulatory processes rather than how often each is reported. The findings reveal three critical structural insights. First, no study taught learners the criteria required to evaluate AI outputs. While interventions frequently prompted learners to pause or explain, none provided evaluative standards, reflecting an implicit assumption that learners already possess adequate evaluation skills. Second, monitoring and evaluation are reported in 94% of studies, and in those reporting metacognitive laziness they are reported together with AI delegation while planning is reported apart from it. Planning is reported in most studies in both outcome groups, so the difference lies in which processes co-occur rather than in whether planning is reported at all. We describe this configuration as verification without criteria. Third, structural configurations vary by intervention timing. Instructor mediation, implemented prior to the first query, anchors delegation within a planning-oriented network, whereas in-task prompting alone isolates delegation from planning. We argue that the placement of friction rather than its amount determines whether verification has a standard, and that evaluative criteria should be established before the first AI query, a proposition this corpus has not tested.