Metacognition

Metacognition — thinking about one's own thinking — is both a target of AI education research (can AI tools develop students' metacognitive skills?) and a risk factor (AI completing tasks may suppress metacognitive practice).(The Evidence Base on AI in K-12: A November 2026 Review)(Designing a mobile chatbot-based learning journaling system for intrinsic motivation and engagement)

Questions to Consider

• 'Metacognition' is thinking about your own thinking — knowing what you know, monitoring yourself, and adjusting your strategies. When you study or solve a problem, how aware are you in the moment of whether you actually understand versus just recognizing the material?
• A striking finding: students who used AI essay assistance were often unable to recall quotes from their own essays, because they hadn't engaged with the content during production. When a tool produces the output, what practice is the learner losing — and is that practice important?
• The page frames metacognition as both a target (can AI build it?) and a risk (can AI suppress it?). Could the same AI tool either strengthen or weaken a learner's metacognition depending on how it's designed or used? What determines which way it goes?
• Structured prompts that ask students to self-explain, evaluate strategies, or identify gaps preserve metacognitive demand, while AI that simply completes tasks displaces it. If you were designing an AI study tool, what would you build so that it invites reflection instead of replacing it?
• The page finds that whether AI use is metacognitively rich depends on the learner's motivation and stance as much as on the technology. Have you ever used a tool in a shallow way and then realized you learned nothing — and what was different about times you used it deeply?

Introduction

Metacognition in education refers to learners' awareness, monitoring, and AI Regulation in Education of their own cognitive processes:

• Metacognitive knowledge: Understanding what one knows, what strategies are available, and when to deploy them
• Metacognitive regulation: Planning, monitoring, and evaluating one's own learning in real time

Within Self-Regulated Learning frameworks, metacognition is the central mechanism that enables learners to adapt strategies, recognize confusion, and seek help appropriately.(Designing a mobile chatbot-based learning journaling system for intrinsic motivation and engagement)

Metacognition is not the same construct as Reflection, and the two are often conflated. Monitoring is a judgment a learner makes about their own state, and it can happen with no record at all. Reflection is the deliberate, documented examination of a completed attempt, and what it produces is a text that can be reviewed or assessed. Keeping them apart matters most when AI is involved: a monitoring judgment is hard to hand over, while a reflection artifact can be produced by a model.

How learners actually deploy metacognition around AI is shaped by more than the tool itself: Cui et al. find that student motivation and the interaction role they adopt shape their metacognitive engagement with GenAI — meaning whether AI use is metacognitively rich depends on the learner's stance as much as on the technology. Miles, Haber-Curran and Arar (2026) add the AI interaction itself as an object of that reflection: the closing step of their Prompt Literacy Cycle asks learners to examine what the process revealed about how prompts function and what assumptions shaped the response, and they make reflection and revision the phase in which authorship and critical judgment develop.

A more fundamental role for metacognition appears when the locus of coordination shifts from teacher to learner. Dang, Hong, Doyle and Nguyen (2026) argue in their Human-AI Collaboration Competency (HACC) framework that the integrative work Technological Pedagogical Content Knowledge (TPACK) assigns to pedagogy passes, once the actor is the learner, to the learner's own metacognition — the competency that lets students direct the collaboration, calibrate their reliance on AI, and remain the authors of their learning. They treat metacognitive competency as the mechanism through which learner agency is enacted, deliberately excluding agency as a separate component and reading it instead as the emergent quality of coordinating domain, AI, and metacognitive competency. Their Epistemic Network Analysis of interviews with 24 higher education educators found metacognitive talk rising with role level — senior educators connected it to domain, AI, and learning with AI — which the authors read as educators with broader curricular responsibility seeing students' persistent difficulties as regulatory rather than technical.

How AI Tools Affect Metacognition

The Suppression Risk (Stanford SCALE, 2026)

When AI completes reasoning tasks for students — solving math problems, writing essays, generating code — the student loses practice in monitoring their own understanding and selecting strategies.(The Evidence Base on AI in K-12: A November 2026 Review)

Reliance is not failure: Du and Yuan (2026) judge AI-mediated learning by six diagnostic criteria — contestability, recoverability, transfer, traceability, distributed responsibility and epistemic plurality — and distinguish instrumental assistance, which helps produce output, from judgment-bearing assistance, which supplies the standards by which output is judged. Naming the governance question. Lin and Chang (2026) propose metacognitive ownership — the learner's governance of the standards, evaluative judgments and actions an AI contributes to — separating standards appropriation, monitoring-judgment governance and control authorization, and requiring a realistic opportunity to act differently.

Key findings:

• Kosmyna et al. (2025): Students who used AI essay assistance were 83% unable to recall quotes from their own essays, vs. 11% for non-AI users — indicating they did not engage with the content during production.
• Stadler et al. (2024): General-purpose AI reduced cognitive load but produced lower-quality reasoning vs. traditional search, suggesting metacognitive engagement was displaced.
• Lehmann et al. (2025): General AI for programming harmed understanding for low-Prior Knowledge students — the students most in need of metacognitive scaffolding received answers instead.
• Metacognitive laziness. The field names a distinct failure mode in which learners abdicate their metacognitive responsibilities to the tool — adopting the term metacognitive laziness for the resulting illusion of competence, and warning that its cognitive risks fall hardest on novices and weaker self-regulators (Lodge & Loble (2026)).

Fluency is why the loss goes unnoticed. Learners judge their learning partly from how easily material is processed, and Bjork et al. (2013) show that fluency cues lead people to overestimate their understanding when content feels easy; fluent AI-generated text supplies that feeling whether or not comprehension followed, creating the conditions for a manufactured illusion of competence. Generativism treats this as the reason adaptive metacognition cannot be assumed: the learner has to ask whether ease of reading reflects understanding rather than reading the ease as the answer.

Awareness does not guarantee that completion stops. Jia (2026) names metacognitively discordant completion: a learner who invested real effort, holds a formed first-person verdict that understanding has not arrived, and releases the work anyway — the case the fluency-illusion and withdrawn-effort accounts both assume away.

Learners are also miscalibrated about AI's costs: participants predicted independent times accurately but underestimated AI-assisted time, and reported lower effort (0.61 points on NASA-TLX) while AI sped up only three of 24 tasks — a bias absent when the imagined helper was another person (Yu et al. (2026)).

The cost appears bounded rather than global: offloading degrades the specific skill practiced, not the underlying domain-general abilities, and surfaces on withdrawal: an endoscopy study found adenoma detection fall from 28.4% to 22.4% when the AI was unavailable, leaving source-monitoring (telling an AI's reasoning from one's own) an open metacognitive risk (Cash et al. (2026)).

The Augmentation Opportunity (Scheu et al., 2026)

When AI is designed to support reflection rather than replace it, metacognition can be strengthened:

• Learning journals are a classic metacognitive practice: by reflecting on learning processes, students increase awareness of their cognition

• Structured prompts that ask students to self-explain, evaluate strategies, or identify knowledge gaps preserve metacognitive demand. CoMeT (Hou et al. 2026) gives that phrase a definition and an empirical warrant: it treats metacognitive demand as a quantity distinct from cognitive load — what the learner must decide, state, or judge before help arrives, not simply what remains when help is withheld — and held it statistically equivalent to a tutor that withheld answers by design (p_TOST = .004) while its own support escalated and faded one rung at a time. Fading held when the learner's turn was aimed at the decision under support: turns aimed elsewhere drew a later concession 40.3% of the time against 28.8% for aimed turns, an 11.5-point difference, so what a tutor must read for is where the learner's attention sits rather than how much effort the turn displays.

• Assign an AI-use level per task. HAIML pairs its three layers with Four AI Use Levels — from no AI use, through brainstorming, to AI collaboration and AI-integrated creation with output evaluation — so an instructor sets the level for a task and keeps metacognitive reflection central at every level (Reardon, 2026).

• The example-based course in Scheu et al.'s chatbot increased perceived competence (a metacognitive Self-Assessment) even when the Large Language Models (LLMs) assistant alone did not

• An emerging framing puts the system, not just the individual, in the loop. A bibliometric–systematic review of 135 STEM-classroom studies (24 core studies) finds a posthumanist framing emerging that positions AI as a co-regulator of learning, shifting the unit of analysis from individual reflection to system-level AI Regulation in Education and distributed cognition (Tsakeni et al. (2025)).

Marrone et al. (2026) fine-tuned GPT-4o-mini to identify four metacognitive sub-elements — Goal Setting, Strategy Choice, Reflection on Learning and Effort Regulation — at the sentence level in 14,100 classroom reflections. Human reliability set the ceiling: initial double coding agreed at κ = 0.32 to 0.51, calibration lifted agreement to κ = 0.52 to 1.00, and model accuracy followed the same ordering — Goal Setting at 91.28%, Reflection on Learning at 67.44%, Effort Regulation at 60.74% — across 18,883 labeled samples.

At synthesis scale the pattern holds, and it thins as measurement sharpens. A three-level meta-analysis of 54 (quasi-)experimental studies (121 effect sizes, 7,431 learners) put AI's effect on metacognition at g = 0.624 (95% CI [0.431, 0.817]), with 107 of 121 effects positive and gains rising with exposure from g = 0.381 under a week to g = 1.638 at 15 to 20 weeks (Li et al. (2026)). Measurement source was the decisive moderator: self-report g = 0.776, behavioral observation g = 0.514, and accuracy ratings, the gap between judged and actual performance, g = 0.035 and non-significant. The gain is largest where it is easiest to claim and near zero where monitoring is scored against performance, which is the self-report problem the studies above keep running into.

Metacognition is among the least-studied outcomes. In a scoping review of 394 GenAI higher-education studies, metacognition and self-regulated learning were among the outcomes targeted least, alongside Ethics and Creativity (Lin et al. (2026)).

• Surfacing interaction patterns that learners cannot see. Kuang, Li and Weng (2026) tracked eye movements, note-taking and speech while 22 interpreting trainees worked with a speech-recognition and machine-translation system, and found that the way students divided attention between AI output and their own notes was invisible to them: 58.3% changed profile between task stages, and the heaviest readers of AI output scored lowest on delivery fluency and target language quality. The pedagogical consequence is that reflection has to be scaffolded by external evidence, because a learner's strategy is not introspectable — the authors argue for guiding students to describe and evaluate why they worked a given way at each stage.

• Process evidence is channel-conditional. In 79 undergraduates working with generative AI, higher-depth monitoring and planning appeared more often in think-aloud units while higher-depth regulation appeared more often in interaction logs, so one record alone understates some processes (Li and Liu (2026)).

• Reflective practices can emerge from ordinary subject-matter use. Across eight Danish upper secondary schools, students working with Generative AI as a cognitive partner developed four practices — explanation, creation, inspiration, and improvement, unfolding into eleven processes — in which reflection appeared in both disciplinary and process registers, though each practice could equally serve substitution (Dalsgaard & Prilop, 2026).

• A reflective scaffold can habituate and replace the self-monitoring it protects. de Jong (2026) warns that repeated prompts to reflect may wear off through habituation — much as repeated warnings are dismissed automatically — so a professional who reflects only when the system asks may end with neither the tool nor the self-monitoring habit; the paper argues reflective prompts should be treated as a fading skill-building intervention and tested for whether users reflect without them.

• A selective gain: metacognitive knowledge without planning and monitoring. Wang et al. (2026) split 68 preservice teachers into a group that explained the flipped classroom to a generative AI novice learner (n=33) and a group that questioned a generative AI teacher (n=35). The novice-learner group scored higher on metacognitive knowledge and learning strategies (p < 0.001, Cohen's d = 0.41) but not on planning and monitoring (M = 4.01 vs. 3.78, p = 0.062), even though the same group's explanation and application scores rose (r = 0.474 and r = 0.642) while factual recall did not separate the groups (p = 0.416). The split follows the two components in the Introduction: explaining to someone who asks back built declarative awareness of what one knows, but nothing in the design made the learner set a route or check progress, which the authors attribute to task design and cognitive load rather than to the role itself. Learning by Teaching with AI is therefore a lever on metacognitive knowledge, not automatically on regulation, and the authors recommend adding explicit planning prompts and progress feedback.

• Think Aloud supports early awareness but not deep evaluation. Karaki and Baroudi (2026) ran a mixed-methods action-research study with 44 first-year entrepreneurship students using the Think Aloud method during AI-supported coursework. Think Aloud made students' prompt construction, verification practices, and ethical decision-making more visible, but critical evaluation of AI outputs "often remained at a surface level, particularly among students requiring additional support" — students spotted generic or misaligned responses yet kept relying on them heavily. Self-reported metacognitive awareness was the highest of three surveyed constructs (M = 3.79), but the single post-intervention survey cannot separate perception from change. The authors' conclusion — Think Aloud is not a stand-alone solution and must be combined with explicit instruction in AI literacy, prompt construction, and verification strategies — extends the process-evidence line above from what records capture to what vocalization alone changes.

• Prior metacognitive training gates the benefit of an AI reflection partner. In a 2 × 2 experiment with 60 pre-service teachers, Zhang et al. (2026) found a significant metacognition × tool interaction: an AI Socratic partner beat traditional digital tools by 0.56 points only after training, with no advantage without it (ΔM = −0.03).

Braun (2026) gives the suppression account a preregistered causal test (N = 317) against worksheets. Disorganization rose by 0.84 (95% credible interval 0.47 to 1.21) and was the clearest effect in the model, while deep processing fell through a chain running from offloading to critical thinking rather than directly. The author names the state cognitive dormancy, in which monitoring and strategy selection go quiet while processing continues, and locates its cause in the affordances of the environment rather than in the learner.

Monitoring changes the route more than it blocks the path. Guo and Dong (2026) split 600 undergraduates by a monitoring item and estimated each group separately: load predicted cognitive offloading at β = 0.844 for low monitors and β = 0.840 for high, effectively the same. Offloading predicted overreliance at 0.396 and 0.506, and the direct load path ran 0.560 and 0.419. Dependence among stronger monitors therefore traveled more through offloading and less directly. These are separate models, not a formal multi-group test.

When monitoring is skipped

Xiao et al. (2026) turned the suppression risk into a measured quantity across four experiments with 728 adults. Higher AI assistance raised task-level disengagement to 4.82 from 3.96 (d = 0.67), and offloading carried 58.0% of the effect (indirect 0.51, 95% CI [0.32, 0.72]). Expert raters then scored the delegated proposals lower: quality fell from 4.78 to 3.95 (d = 0.75), with 57.8% of that drop running through disengagement (indirect -0.48, 95% CI [-0.68, -0.30]).

The cost is conditional. Task complexity strengthened that path (index -0.39, 95% CI [-0.62, -0.16]) while AI literacy weakened it (index 0.30, 95% CI [0.09, 0.52]). So monitoring is not only a capacity that predicts better use; skipping it is a step whose absence is visible in output quality through an offloading route. Two limits bound the claim. The AI was a researcher-scripted response bank, not a live model, and occupation and student status went unreported, so this is neither a classroom nor a workplace study.

The Engagement–Motivation Distinction

Scheu et al. (2026) found a critical split:

This suggests that metacognitive support and Motivation are not identical. The LLM assistant's Scaffolding of journal entries increased how much students wrote (behavioral engagement) but did not make them want to write more (intrinsic motivation).(Designing a mobile chatbot-based learning journaling system for intrinsic motivation and engagement)

The Beliefs-vs-Experiences Distinction

Guo & Ye (2026) offer a theoretically sharper account of how metacognition governs strategy selection, distinguishing two components that operate in different phases:

• Metacognitive beliefs — stable, self-referential self-conceptions stored in long-term memory (e.g., beliefs about one's memory capability, or the reliability of a tool). These anchor strategy choices before task initiation.
• Metacognitive experiences — dynamic, task-specific feelings (perceived difficulty, confidence, mental workload) that drive belief updating during task execution.

This distinction yields the principle of timing-component matching: feedback that targets beliefs (e.g., comparative rankings) is most effective in the pre-task preparation phase, whereas feedback that targets experiences (e.g., immediate correctness indicators) is most effective during task execution. Abstract ranking feedback can become separated from — or overridden by — the task-specific experiences that dominate immediate decision-making, explaining why some feedback interventions fail to change behavior. This gives educators a phase-contingent rationale for designing metacognitive scaffolds around AI tools: calibrate beliefs before use, provide immediate task-specific feedback during use.

External support can also step in front of the judgment entirely. Iqbal et al. (2026) had 87 EFL students revise an essay with ChatGPT 4.0, an experienced academic writing instructor, or no support, and reconstructed the strategies students used from clickstream, keystroke and mouse traces. What they chose was strongly associated with the support available (Cramér's V = 0.668), but not with metacognitive judgment accuracy (H(3) = 4.995, p = 0.172), which showed a marginal link only in the no-support condition (p = 0.0460). No revision strategy was associated with essay-score change (p = 0.273), yet the GenAI group improved significantly more than the other two (H(2) = 16.591, p = 0.00025, η² = 0.174), so the stronger essays did not come from better self-regulation. When a capable partner can supply the next move, metacognitive judgment stops being what selects it, and the strategy-choice account above describes the unaided learner more than the supported one. The authors' practical reading is to aim AI assistance at prompting reflection on writing strategies rather than at supplying direct help.

Calibration is trainable: prediction + feedback

Ngai & Gilbert (2026) provide direct causal evidence that metacognitive calibration is a trainable skill. In two preregistered experiments (N=164, N=416), just five practice trials pairing a performance prediction with veridical feedback improved calibration and reduced bias. A four-group additive design isolated the causal component: making predictions alone was ineffective; adding performance feedback drove the improvement; explicitly labeling over-/underconfidence added nothing further. Critically, the improvement acted on absolute calibration — raising confidence in the underconfident and lowering it in the overconfident — so it corrected miscalibration in both directions rather than shifting everyone one way (which is why signed/directional effects were null). This strengthens the "experiences not beliefs" account above and shows the minimum viable metacognitive training: prediction + immediate, task-specific feedback.

• A brief reflection prompt sharpens monitoring during AI-supported decisions. Ren (2026) added three reflection prompts before finalizing answers in a three-condition experiment with 342 undergraduates: acceptance of incorrect ChatGPT advice fell from 62.4% to 39.7% (OR = 0.40) and awareness calibration rose (0.59 vs. 0.41), while recommendation accuracy and alignment with correct advice stayed high. Reflection made reliance more discriminative rather than uniformly defensive, which supports treating reliance as a monitoring problem rather than a question of how much AI is used.

• Nest prompting inside a metacognitive cycle. The IDEA framework sequences GenAI use as Intent, Deconstruction, Expression, and Adaptation, making prompting one step in planning and monitoring rather than a stand-alone skill; students taught it produced higher-quality prompts than peers taught Role–Task–Context–Format prompting in all five task categories (Wang et al., 2026).

Teacher accounts of metacognition need the same scrutiny as student self-reports. In 25 Chinese university teachers, a shared boundary around what AI cannot do — emotion, embodiment, metacognition, tacit knowledge, judgment — appeared in every case, and verification carried clear evidence in 24 of 25. But both codes came after a direct question about what AI cannot replace, so the agreement describes how teachers answered that question rather than a concern they raised unprompted. On the intellectual-degeneration proposition the protocol offered, nine endorsed it and eleven made it conditional. The finding is a measurement caution: a belief that looks universal under prompting may be an artifact of the prompt (Qiang & Wang, 2026).

Implications for Tool Design

1. Preserve the "friction" of thinking: If AI writes the reflection, the student does not build metacognitive skill. Journaling assistants should scaffold, not author.
2. Model metacognitive language: The example-based course worked partly because it exposed students to proficient models' metacognitive self-talk.
3. Separate support for motivation vs. skill: Metacognitive skill development (course-structured) and productivity enhancement (AI-assisted) may require different design strategies.
4. Name the target as onloading, and ration the help. A metacognitive approach to learning and performance in human-AI interaction separates cognitive offloading, which stays under metacognitive control, from metacognitive offloading, where planning, monitoring, evaluation and judgment calibration are handed over. Their design target is metacognitive onloading: scaffolds that make the learner, not the AI, set goals, anticipate difficulty and judge strategies, placed within a prompt, across a task, and fading over time. The help rule is minimum-necessary: a hint or diagnostic question first, with deeper assistance unlocked only after the learner asks, monitors or chooses a direction. The paper is conceptual and its authors note that no validated measure of either construct yet exists.

AI may alter the metacognitive threshold for deciding one knows enough to answer: Marcoccia et al. (2026) found that mere access to AI advice suppressed people's willingness to suspend judgment under uncertainty, even with wrong advice and accuracy incentives — an effect that survived unsolicited AI output and monetary stakes.

Proactive agentic AI can displace the learner's own metacognitive loop: Woollaston et al. (2026) argue that when agents pre-fetch, initiate, and self-correct, the agent's planning, monitoring, and evaluation replace the learner's, removing the retrieval practice and self-monitoring that desirable difficulties and metacognitive training depend on.

• Mistake-based Pedagogies and Teaching Strategies as metacognitive training: Hosseini (2026) shows that deliberately exposing students to AI-generated errors in a database design course activates metacognitive monitoring — students inspected outputs, identified errors, and revised designs rather than accepting AI output at face value. Self-reported AI literacy correlated weakly and negatively with objective competency (r=−0.39), a calibration gap the critique-refinement cycle is designed to narrow.

• Productive failure engages metacognitive monitoring. Kim et al. (2026) show productive-failure-based learning activates reflection on one's own attempts; clue-before-correction tasks require learners to diagnose and correct their own errors — a metacognitive activity where AI gives clues rather than answers.

• Self-regulation buffers offloading harm but cannot cancel it. Chen (2026) shows that metacognitive regulation (self-regulated writing) attenuates the negative association between deep cognitive offloading and independent no-AI outcomes in GenAI-assisted writing (interaction B = 0.22), but does not eliminate it — a bounded-support condition pairing delegation limits with compulsory reflection about how AI suggestions were accepted/rejected produced the strongest independent performance.

• Explanation-seeking depth predicts task quality, not recall. Tsiligkiris (2026) shows explanation-seeking prompts (depth) in LLM interaction predict task quality but not immediate recall, interpreting the dissociation via elaboration (comprehension) vs. retrieval practice (consolidation) — and suggesting explanation-seeking correlates with metacognitive monitoring, though retrieval demands must be added for durable retention.

• Self-reported metacognition is a weak proxy for regulation with an LLM. Clerc et al. (2026) gave 116 middle-school students a two-hour AI literacy workshop and then measured their LLM interaction during science problems: trained students accepted underspecified prompts less often (51.5% vs. 66.7%), asked follow-up questions after a weak response far more often (59.2% vs. 27.9%, d = 0.80) and judged answer correctness more sensitively to prompt quality (interaction OR = 2.52). Neither a general metacognitive-awareness scale (Jr. MAI) nor GenAI self-reports predicted those behaviors or final performance (r = .04 and r = .01) — monitoring and control during generative AI use is task-specific, and observable behavior carries more information than the self-report instruments built to capture it.

• Automated scoring reached the performance phase, not the planning that precedes it. Chen and Liu (2026) gave 46 interpreting students 14 weeks of weekly automated scoring with a returned score, marked errors, and a reference rendition: the automated group gained more overall (d = 1.03). Only monitoring during practice correlated with score gains (r = 0.42), while pre-learning planning sat near the scale midpoint (M = 3.01). Evaluation and reflection carried the second-highest mean (3.87 of 5) yet showed a near-zero link to gains (r = 0.10), so a high reflection score should not be read as productive reflection.

• Verification literacy pays off only through metacognitive self-regulation. Davor, Larbi and Boateng (2026) surveyed 533 university students and found that AI verification literacy had no direct association with critical thinking or technical Problem Solving; it mattered only indirectly, through metacognitive self-regulation (a full mediation pattern). Cognitive offloading tendency ran the other way, predicting lower self-regulation (-.294) along with lower critical thinking (-.240) and problem-solving (-.312).

• Reflection rose with AI access, and it fed reasoning rather than test scores. Melanou, Beege and Kimmig (2026) followed three classes (N = 87) across a nine-week course, measuring at three points. Knowledge improved in all conditions with no advantage for either AI condition and no Matthew effect (BF01 = 8.70), but reflective use (checking sources and verifying AI output before adopting it) was far higher in the AI condition than the control group (M = 3.72 vs. 2.82, F(1, 40) = 20.21, p < 0.001) and predicted critical thinking (β = 0.43, p < 0.001, R² = 0.183) while predicting no knowledge gain. Reflection appeared from tool access alone rather than from scaffolding, and because its payoff lands in reasoning, a knowledge test will not register the metacognitive benefit of an AI-supported course.

• Inside a metacognitive self-report instrument, awareness outpredicts the rest. Alatoai and Alshahri (2026) validated a 45-item measure of AI-supported STEM learning with 649 Saudi secondary students (CFI = 0.983, RMSEA = 0.019; subscale ω = 0.888–0.905) and used it to rank the components. The four dimensions explained 68% of the variance in self-regulated learning scores, with AI-based metacognitive awareness strongest (β = 0.38, p < 0.001), cognitive transfer and adaptability next (β = 0.29, p = 0.008), and creative and critical AI-STEM reasoning not significant (β = 0.14, p = 0.135). Awareness is the component that carries adaptive performance, the authors advise reporting subscales separately rather than one global score, and they reserve higher-order reasoning for reasoning-quality tasks; the measure is nonetheless self-report, so like the workshop study above it captures perceived capability rather than observed strategy use.

• The evaluator's own monitoring is metacognitive work too. Hoppe, Loibl and Leuders (2026) argue that an AI-generated diagnostic inference is not raw evidence but an interpretation already made, so teachers must integrate it with their own observations in a process they call meta-diagnosis, deciding deliberately whether to accept, reject, or modify it. That places a second metacognitive loop beside the learner's: not only how the student regulates thinking with AI, but how the teacher evaluates what the system claims about that thinking.

Common Misconceptions

• "AI assistance leaves the underlying ability intact." Students who used AI essay assistance were 83% unable to recall quotes from their own essays, against 11% for non-AI users, indicating they did not engage with the content during production. The production process is where the engagement happens, and bypassing it shows up later.
• "Offloading degrades everything." The cost appears bounded rather than global: offloading degrades the specific skill practiced, not the underlying domain-general abilities, and surfaces on with-something transfer tests. The localisation matters, because it says what to protect rather than warning about general decline.
• "Ability to produce implies ability to retain." The 83% versus 11% recall gap is the direct counterexample: output was produced in both conditions and the memory trace differed enormously.

Connected Concepts

• Pedagogical Patterns — The reflection steps these sequences build in
• Learners — Learners: the umbrella for the learner-side concepts
• Self-Regulated Learning
• Self-Assessment
• Cognitive Offloading
• Scaffolding
• Agentic AI
• Formative Assessment
• AI Literacy
• Retrieval, Spacing and Interleaving — judgments of learning and the fluency illusion that retrieval practice corrects
• Cognitive Surrender
• Reflection

Connected Articles

• Teaching students to question the machine: An AI literacy intervention improves students' regulation of LLM use in a science task — a two-hour AI literacy workshop shifted middle-school students' LLM-interaction regulation, unlike their self-reported metacognition (Clerc et al. 2026)

• Student-AI interaction in computer-assisted consecutive interpreting: patterns and performance — Student-AI Interaction in Computer-Assisted Consecutive Interpreting

• Epistemic dependence in AI-mediated learning — Epistemic dependence in AI-mediated learning (Du & Yuan 2026)

• What students ask matters: LLM interaction depth, task quality, and immediate recall in higher education — What students ask matters: LLM interaction depth, task quality, and immediate recall (Tsiligkiris 2026)

• Layer-sensitive cognitive offloading in generative AI-assisted writing: Supported performance and independent no-AI outcomes — Layer-sensitive cognitive offloading in GenAI-assisted writing (Chen 2026)

• Efficiency vs. Effectiveness: Self-Regulated Learning with LLM-Mediated Help-Seeking — LLM-mediated help-seeking in STEM: layered, instrumental, and verified

• How motivation and roles influence metacognitive engagement in student-GenAI interaction — Motivation and roles in metacognitive GenAI engagement

• Metacognitive training facilitates optimal cognitive offloading — Metacognitive training facilitates optimal cognitive offloading (Ngai & Gilbert 2026)

• Meta-cognitive insights into cognitive offloading: Mechanisms, interventions, and educational implications — Meta-cognitive insights into cognitive offloading: mechanisms, interventions, and educational implications (Guo & Ye 2026)

• The IDEA Framework for Metacognitively Regulated GenAI Use in Higher Education: Development and Exploratory Pilot Evidence — The IDEA framework for metacognitively regulated GenAI use

• Introducing HAIML: A Human-Centered AI Metacognitive Learning Model — A Framework for Human Agency and Reflective Learning — HAIML: a human-centered AI metacognitive learning model (agency & reflective learning)

• Metacognitively discordant completion and the aware pass-through of non-understanding in generative AI learning — Metacognitively discordant completion and aware pass-through of non-understanding

• Mapping the Scaffolding of Metacognition and Learning by AI Tools in STEM Classrooms: A Bibliometric–Systematic Review Approach (2005–2025) — AI tools scaffolding metacognition in STEM

• Is AI making us stupid — Is AI making us stupid? critique of cognitive offloading

• The Evidence Base on AI in K-12: A November 2026 Review — General-purpose AI suppresses metacognition by completing reasoning

• AI advice suppresses people's willingness to say "I don't know", even when the advice is wrong and accuracy is incentivized — AI advice suppresses willingness to say "I don't know", even with wrong advice and accuracy incentives

• Agentic AI and Pedagogical Best Practice: The Tension Between Automation and Learning — Agentic AI and pedagogical best practice: the tension between automation and learning

• Cognitive offloading and the speedup illusion in human-AI interaction — Cognitive offloading and the speedup illusion in human-AI interaction

• Artificial intelligence, cognitive offloading and implications for education — AI, cognitive offloading and implications for education (Lodge & Loble 2026)

• The Pedagogy of AI Mistakes: Fostering Higher-Order Thinking — The Pedagogy of AI Mistakes: Fostering Higher-Order Thinking (Hosseini 2026)

• Designing AI systems to support a productive-failure-based learning: insights from adult learners on AI applications and AI system design principles — Designing AI Systems to Support Productive-Failure-Based Learning

• Clue before correction: ChatGPT-enhanced strategy for promoting autonomous and reflective language learning — Clue Before Correction: ChatGPT for Autonomous Language Learning

• Prompts to Practice: A Pedagogical Framework for Human-Centered AI Engagement — Reflection on the prompting process and authorship development in the Prompt Literacy Cycle (Miles, Haber-Curran & Arar 2026)

• College students’ metacognitive awareness of generative-AI reliance: an experimental study of decision confidence and attribution bias — A reflection prompt cut acceptance of incorrect AI advice and improved awareness calibration (Ren 2026)

• Impact of automated scoring on interpreting performance and self-regulated learning: evidence from a pedagogical experiment — Automated scoring reinforced monitoring but not planning, and self-reported reflection stayed unproductive (Chen & Liu 2026)

• Artificial intelligence-supported learning and higher-order cognitive outcomes: the mediating role of metacognitive self-regulation — Verification literacy acting only through metacognitive self-regulation (Davor, Larbi & Boateng 2026)

• Human or GenAI Support? Conditions Impacting Students' Strategy Choices in an Essay Revision Task — External support, not metacognitive judgment accuracy, drove revision strategy choice (Iqbal et al. 2026)

• The Contribution of Generative Artificial Intelligence as a Novice Learner to Students in the Learning by Teaching Model — A selective gain: metacognitive knowledge rose while planning and monitoring did not (Wang et al. 2026)

• Generative AI and Learning Dynamics in Higher Education: A Longitudinal Empirical Study — Reflective use rose with AI access and predicted critical thinking, not knowledge gain (Melanou, Beege & Kimmig 2026)

• Exploring the Impact of AI-Based Learning Environments on Student Self-Regulation and Adaptive STEM Learning — Metacognitive awareness as the strongest predictor of adaptive STEM learning (Alatoai & Alshahri 2026)

• Rethinking teachers’ diagnostic skills in AI-supported formative assessment: from diagnosis to meta-diagnosis — From diagnosis to meta-diagnosis: teachers judging AI-generated inferences (Hoppe, Loibl & Leuders 2026)

• Open Questions Towards Skill-Sustaining Reliance in Reflective AI Engagement — Open Questions Towards Skill-Sustaining Reliance in Reflective AI Engagement

• Conceptualising Human–AI Collaboration Competency for hybrid intelligence: towards learner agency through a metacognitive perspective — Metacognition, not pedagogy, as the mechanism of learner agency in human-AI collaboration (Dang, Hong, Doyle & Nguyen 2026)

• Metacognitive Ownership in Human-AI Regulation: Construct Definition, Boundaries, and a Research Agenda — Metacognitive ownership: a construct for learner governance of standards, judgments and actions

• Mapping the Distribution and Depth of Metacognitive Processes in Generative AI-Assisted Learning: Evidence from Interaction Logs and Concurrent Think-Aloud Protocols — Logs and think-aloud capture different metacognitive processes in GenAI-assisted learning

• AI as a cognitive partner in learning: Emerging epistemic practices of reflective thought — four practices and eleven processes of reflective thought when students treat AI as a cognitive partner

• Ethical generative AI use in higher education: An action research study using the Think Aloud method — Think Aloud made prompt construction, verification and ethical reasoning visible, but critical evaluation stayed surface-level; not a stand-alone solution

• Metacognitive training and AI-based Socratic scaffolding in pre-service teachers' reflective practice: a dual-study experimental design — Metacognitive training and AI-based Socratic scaffolding in pre-service teachers' reflective practice: a dual-study experimental design

• Cognitive dormancy: Supplanting structured learning with AI affordances — Cognitive dormancy: an experimental account of suspended metacognitive engagement and rising disorganization (Braun 2026)

• Can artificial intelligence promote metacognition? Evidence from a three-level meta-analysis — Three-level meta-analysis: pooled g = 0.624, falling to non-significance for accuracy-rated monitoring (Li et al. 2026)

• A metacognitive approach to learning and performance in human-AI interaction — SYNC model: metacognitive offloading vs. onloading, and task performance decoupled from durable learning (Fan et al. 2026)

• University students’ overreliance on generative AI feedback in the learning process: the effects of perceived cognitive load, cognitive offloading, and metacognitive monitoring — Monitoring does not block the offloading path so much as reroute it

• Metacognitive disengagement and knowledge innovation quality in human–AI collaboration: the mediating role of cognitive offloading and the moderating roles of task complexity and AI literacy — When monitoring is skipped, and what it costs