Metacognitively Discordant Completion and the Aware Pass-Through of Non-Understanding in Generative AI Learning

Synthesis: This theoretical paper names a state it calls metacognitively discordant completion (MDC): a learner submits correct, complete work while holding a first-person awareness that understanding has not actually arrived. Arguing that no existing literature holds the three defining conditions together under one name, the author builds the construct by inheritance from Metacognition research and by boundary against related concepts, framing GenAI's role as amplification rather than invention.

Key Findings

1. The construct. MDC is the experiential state of a learner who has invested cognitive effort, holds a formed first-person awareness that understanding has not occurred (judged against their own operative standard), and releases the completion anyway. The conjunction of the three conditions is the construct.
2. Not new to AI. A pre-AI genealogy across five literatures establishes that the dissociation between completing and understanding is old; four cuts separate MDC from withdrawn effort, justification discourse, fluency illusions, and preprint-stage constructs.
3. Sibling construct and boundary. A sibling construct, the Absent Cognitive Baseline, marks the case where no verdict can form; frameworks from epistemic akrasia to academic dishonesty each lack one element of MDC. GenAI's part is amplification, not invention.

The Construct in Three Conditions

Metacognitively discordant completion names the state of a learner who has invested real cognitive effort in an academic task, who holds a formed, first-person awareness that understanding has not occurred — judged against the learner's own operative standard — and who releases the completion anyway, submitting the work and moving on. Condition (i) requires that effort went into the task, separating the state from every account built on withdrawn investment. Condition (ii) requires that the awareness of non-understanding is formed and first-person, a verdict held at the moment of completion, separating the state from accounts built on being deceived by fluency. Condition (iii) requires that completion is released in the presence of that verdict. The paper argues no existing literature holds all three together under one name at one moment.

The author opens with a vivid case: a learner in an introductory statistics course who works through a violin-plot interpretation with a generative assistant, produces a correct and well-organized write-up, and would say in the same breath that the reading is right and that she could not reproduce it on a new chart — because the understanding never arrived. She is not deceived about her state, broke no rule, and invested real effort. The miss in existing vocabulary is systematic: the illusion family assumes the learner believes understanding happened when it did not; the integrity family assumes a rule was broken; the engagement family assumes effort was withdrawn. MDC rests on none of these.

Inheritance from Metacognition Research

The construct is grounded in the Metacognition literature. The paper draws on metacognitive experiences and the learner's own standard, on the two-level monitoring/control framework in which a verdict is formed above and action follows below, and on comprehension monitoring research showing how easily failures of understanding go unnoticed. Two findings anchor condition (iii). First, judgments of learning are causally connected to study choice — people act on them — so a completion released against a standing negative verdict is worth singling out. Second, the traditions that model the decision to answer, from free-report withholding to the Diminishing Criterion Model, run entirely on one signal: confidence that the answer is right. They never needed a second signal because the answer was the person's own work, so confidence in the answer and confidence in one's own understanding moved together.

Generative AI takes that simplification apart. A learner can hold high confidence that the submitted work is correct — because the system that produced it is usually right — while holding a clear negative verdict on their own understanding. The release decision reads the first signal, which approves. Nothing in the judgment economy fails; the criterion is met and the answer is good, but the completion passes through carrying a verdict these models have no place for.

A Pre-AI Genealogy

Five literatures documented the dissociation before generative AI existed. The learning-versus-performance tradition established that performance during acquisition and durable learning are dissociable, but never enters the learner's registration of the split as a variable. The surface-and-deep approaches tradition recorded students who could summarize but not demonstrate understanding, yet an approach is a cross-situational orientation written from the demand side. Cognitive Offloading research defines offloading as using physical action to alter the information-processing requirements of a task, with memorial consequences for externally available information, but no element of the framework occupies the moment of completion. The intelligent-tutoring tradition named gaming the system — behavior aimed at obtaining correct answers by exploiting software with no need to understand why — yet its harm is defined by pre-post outcomes, never by anything the learner reports. And the comprehension-monitoring tradition documented detection that stops short of repair, but its frame is deficit and repair, not deliberate release.

Each of these literatures predates the tool, which licenses the amplification claim: a state on record before the tool cannot be the tool's invention. The author notes that everyday language has long named the phenomenon — a Chinese verb for delivering work to discharge an assignment, a saying for knowing that something is so without knowing why, and the German Bulimielernen — the usual signature of a construct gap rather than redundancy.

Boundary Cuts Against Neighboring Concepts

The paper removes four vocabularies that could absorb the construct. It is not withdrawn effort: metacognitive laziness describes learners handing self-AI Regulation in Education to a generative assistant, and such designs measure regulation through trace data that cannot see an awareness of a comprehension gap at the moment of submission. It is not a justification discourse: MDC is an experienced condition that exists before any reason is assembled, and folding in the neutralization tradition would replace the state with the story told about it. It is not a fluency illusion: in every branch of the illusion family the learner does not know, while MDC names the learner who does. It is not captured by adjacent preprint constructs such as comprehension debt or epistemic debt, which describe costs that accumulate without the person noticing and sit on the illusion side of the awareness axis.

The nearest epistemic-akrasia structure does not attach either: the MDC learner holds consistent beliefs and acts on them, so the state is epistemically enkratic rather than akratic. Cognitive dissonance implies a drive to discharge inconsistency, while MDC includes learners who reside in discordance across many completions without reducing it. Distinct lines of engagement research treat disengagement as an intentional choice without requiring a formed verdict, and deliberate ignorance arranges not to come to know, whereas the MDC learner already possesses the negative verdict.

The Border with the Absent Cognitive Baseline

The border case is a sibling construct. The Absent Cognitive Baseline names the case in which a tool did the work before the skill ever formed, so the learner has no internal standard against which a verdict about understanding could issue at all. MDC names the opposite: the baseline exists, the learner uses it, the reading comes back short, and the completion goes out anyway. One learner has no ruler; the other holds the ruler, lays it against the work, reads the shortfall, and lets the work go all the same. The border is not just wording, because two constructs that sort the same case differently — the ACB learner cannot say whether a comprehension check was passed, while the MDC learner says it was not — are not one construct with two names. Condition (ii) also involves an awareness gradient, from a faint unease to a recognition the learner could put into words, and release may divide into deferred and closed subtypes.

What this means for practice

• Instructors. Stop reading a correct, well-formed submission as evidence of understanding: the paper's argument is that whenever this state occurs, an Assessment that treats completion as evidence of understanding is measuring the wrong one of the two, structurally rather than by occasional error.
• Instructors. Build one transfer check into each AI-permitted task — a new chart, a new dataset, a fresh problem — so a learner who cannot reproduce the reading must surface it, as the introductory-statistics case in the paper could not reproduce the violin-plot interpretation on a new chart.
• Instructors. Handle the state as a cognitive condition rather than a conduct case: the MDC learner invested real effort, broke no rule, and was not deceived, so detection and enforcement do not reach it and the Academic Integrity frame misses one of its defining elements.
• Learners. Name the moment of release: when you submit work that is correct but that you could not reproduce, write down the verdict you were holding at that point, so a deferred completion can be revisited rather than closed.
• Researchers. Design the evidence before prescribing the remedy — an account counts as evidence for MDC only when completion, detected non-understanding, and release are held together at one moment; a learner who believed understanding was present and found out later is evidence for the illusion family or the Absent Cognitive Baseline, not for MDC.

Limitations

• The work is conceptual and reports no data: the author states the construct's status is conceptual and assigns every empirical question to a planned interpretative phenomenological study using interviews and reflective journals with the GenAI-concurrent cohort.
• Graphicacy and mathematics are the expected empirical domains, not boundaries of the construct, so nothing yet shows that the state arises under the same conditions in other subjects.
• The lower edge of the awareness gradient is undecided — where a faint unease ends and a codable recognition begins is one of four evidence debts the paper explicitly leaves to future coding, alongside whether deferred completions ever convert into understanding.
• The justification layer is excluded by design and developed elsewhere, so the paper deliberately does not account for how reasons are assembled around the state and declines to prescribe any assessment redesign until the assigned evidence arrives.

Citation

Jia, Y. (2026). Metacognitively discordant completion and the aware pass-through of non-understanding in generative AI learning. EdArXiv preprint.