Metacognitive Ownership in Human-AI Regulation: Construct Definition, Boundaries, and a Research Agenda

Synthesis: This conceptual paper introduces metacognitive ownership: learner governance of the standards, evaluative judgments, and actions in a learning episode where AI or another external source contributes to at least one of them. Drawing on self-regulated learning, hybrid human–AI regulation, Learner Agency, and evaluative judgment, the authors define three components — standards appropriation (which criterion actually guided the learner), monitoring-judgment governance (how the learner reached an evaluation), and control authorization (what guided the next action, and whether a realistic alternative existed). A running hypothetical example, Lina, responds to the same AI advice in three ways: confidence-based copying, genuine evaluation, and evaluation that reaches a wrong conclusion. From this, the paper develops four interpretive principles and two tentative propositions about later advice decisions and continued regulation when support is reduced. Correctness, advice acceptance, and formal approval cannot establish ownership; the work is a research agenda whose value depends on whether episode profiles prove reproducible and add information beyond existing constructs. No GenAI tool is evaluated and no empirical data are reported.

Key Findings

1. Metacognitive ownership is learner governance of the standard, evaluation, and action within an episode where an external or jointly produced contribution — including generative AI output — enters at least one function.
2. Its three components are standards appropriation, monitoring-judgment governance, and control authorization; each gets a separate evidence-based conclusion, because governing one function does not establish the others.
3. The same outcome can follow different processes: Lina's correct revision results from confidence-based copying in one version and from evaluating the study design in another — procedural approval versus ownership.
4. Control authorization also requires a realistically usable alternative action and evidence that the action followed the learner's evaluation or an applicable prior delegation policy adopted earlier.
5. Ownership does not guarantee correctness — Lina's third version is learner-governed yet wrong, because she applies a deficient criterion — so competence and ownership must be assessed separately.
6. Four interpretive principles and two tentative propositions follow, linking full-cycle ownership to later selective reliance and calibration, and to learner-directed regulation after support is reduced.

Three components and the evidence each requires

Standards appropriation asks which criterion actually guided the learner's judgment or action. Showing a rubric is not enough: evidence must show the criterion was used, and a rejected AI-proposed criterion must be replaced by an identified alternative. Monitoring-judgment governance asks how the learner reached an evaluation and requires evidence connecting something they examined — a task-relevant feature, reason, or criterion — with the judgment that followed; confidence in the AI, polished wording, and time on screen cannot establish that connection. Control authorization asks what guided the next action and whether another action was realistically usable. Accepting a revision because it is the default or the AI sounds authoritative, without evaluation or an applicable earlier decision, is procedural approval. An earlier instruction authorizes a later action only if the record shows the policy predated the episode, covered the action, and stayed open to withdrawal.

Lina's three responses to the same advice

The case is illustrative, not empirical. Lina must evaluate the claim that an AI tutoring program caused a 15% increase in test scores; an observational pretest-posttest comparison without a comparison group is her evidence. The AI correctly explains in all three versions that maturation, concurrent instruction, testing effects, instrumentation, or regression to the mean could explain the gain, and recommends a cautious statement of association. In the first version, Lina copies it because the system sounds certain — procedural approval despite a correct answer. In the second, she notices the missing comparison group, applies the rubric's evidence-strength criterion, and revises deliberately. In the third, she examines the reasoning but treats a large pretest-posttest gain as sufficient for causality. Versions two and three both support all three components; only one reaches the right conclusion.

Distinguishing ownership from related constructs

The authors locate the framework's value in episode-level connections rather than newly identified functions. SRL and hybrid regulation already describe how standards, monitoring, and control can be distributed between learners and AI, and a recent GenAI-SRL measure of second-language writing regulation already includes governance of tool use. Formative assessment and evaluative judgment cover standards, comparison, and improvement plus the capability to judge quality — but capability and its exercise in a particular decision answer different questions. Ownership of learning is a broader college-readiness model in which Metacognition is one component; Trust and reliance research separates attitude from advice use; and Cognitive Offloading describes reduced processing without saying who directed the choice. If profiles repeat what existing approaches capture, the case for a separate construct weakens.

Two propositions and a research agenda

The agenda asks whether component conclusions are reproducible inside a defined monitoring–control cycle and whether profiles add information beyond a baseline of learner governance, initial performance, competence, general SRL, Trust, and AI use. Researchers could compare repeated, matched tasks with similar outcomes but different decision processes, varying advice correctness and whether it appears before or after an initial learner judgment. Classifications should be made from process evidence before correctness or later outcomes are examined. The first proposition expects a more positive link between full-cycle ownership and later selective reliance or confidence–accuracy calibration when competence is higher; the second expects repeated full-cycle profiles to predict learner-directed regulation on a new related task after support is reduced.

What this means for practice

• Instructors. Separate difficulty directing a decision from difficulty judging the subject matter. Lina's first response is correct but follows copying, so feedback should ask why the study design limits the claim; her third uses an inadequate standard, so teaching should address causal inference.
• Instructors. Build in usable opportunities to inspect a criterion, examine the basis of an evaluation, and retain, revise, or reject a recommendation — within learners' time, language, and access conditions.
• Instructional designers. Show the evidence beside the AI's suggestion: study-design information next to the recommended revision helps a learner judge whether a claim should change, and delegated policies must stay accessible and revocable.

Limitations

• The paper reports no empirical data: its central case is explicitly "not an empirical case," and the authors state that the framework's relationships with learning require empirical evaluation.
• Both propositions are tentative and untested: the paper specifies what evidence would challenge them but offers no estimate, and the studies behind them tested their own designs.
• The scheme depends on activity records and learner explanations that are often incomplete or reconstructed after the event, and group-level governance would require specifying how decisions are shared.

Citation

Lin, M. P.-C., & Chang, D. (2026). Metacognitive Ownership in Human-AI Regulation: Construct Definition, Boundaries, and a Research Agenda. PsyArXiv preprint.