In one room, a student writes a learning journal about how they approached a difficult problem.
In another, a participant presses one of two buttons and rates confidence after hundreds of perceptual trials.
Both studies may be described as metacognition research.
They may barely resemble one another.
The 2021 npj Science of Learning review by Damien S. Fleur, Bert Bredeweg and Wouter van den Bos, “Metacognition: ideas and insights from neuro- and educational sciences,” takes that mismatch seriously.
Its central contribution is not a new definition of metacognition. It is a map of how two major research traditions operationalize the construct differently and what might be gained by connecting them.
The authors are well positioned for that bridge. Their affiliations span the University of Amsterdam, Amsterdam University of Applied Sciences and the Max Planck Institute for Human Development in Berlin. The review is therefore also a strong European anchor for the Institute’s Metacognition & Learning hub.
One word, several research objects
The review begins from a broad distinction familiar to the field: metacognition includes knowledge or awareness of cognition and regulation or control of cognition.
Educational research has often studied these capacities at the level of learning strategies, self-regulation, reflection and academic achievement.
Cognitive neuroscience and experimental psychology more often isolate narrower operations: confidence about individual decisions, error monitoring, executive-function tasks, neural correlates and computational measures.
Neither approach owns the construct.
They create different observational windows.
This is why apparently contradictory findings can sometimes reflect measurement differences rather than true theoretical conflict.
A self-report questionnaire about study habits and a trial-level confidence measure are not expected to converge perfectly because they operate at different timescales and ask about different targets.
Offline and online metacognition
A particularly useful distinction in the review is between what can be called offline and online metacognition.
Offline processes include explicit knowledge about strategies, planning before a task, reflection after performance and broad judgments about one’s own learning. These processes are often accessible to language and can be studied through questionnaires, interviews, journals or prompts.
Online processes unfold during performance: detecting uncertainty, monitoring an error, adjusting control or estimating confidence on a specific trial.
The boundary is not absolute.
Planning can occur during a task. A confidence judgment can draw on long-term self-knowledge. But the distinction helps explain why educational interventions often target reflective and strategic processes while laboratory paradigms focus on tightly timed monitoring.
Measurement traditions create different findings
The review compares methods used across fields.
Educational research commonly uses self-report inventories, interviews, think-aloud protocols, learning journals and metacognitive prompts. These methods can capture strategic knowledge and reflection in ecologically meaningful settings, but they may depend on verbal ability, self-insight and response biases.
Experimental research often uses confidence judgments, objective performance, signal-detection measures and executive-function tasks. These approaches can tightly quantify relationships between performance and monitoring but may have limited ecological validity and uncertain transfer to real-world learning.
The paper reports that self-report questionnaires and interviews tend to be only weakly related to academic achievement, whereas think-aloud protocols and executive-function measures show stronger relationships in the literature the authors review.
That pattern should not be read as “think aloud is true and self-report is false.”
It suggests that measurement proximity, task demands and behavioural anchoring matter.
Academic achievement is not the same as metacognition
Educational studies often use academic performance as an outcome.
That is reasonable, but it creates another interpretive layer.
A learning strategy can improve metacognitive regulation without immediately changing a course grade. A grade can improve because of subject knowledge, motivation, prior attainment, teaching quality or social support rather than metacognition. Executive-function performance may correlate with achievement through pathways that are not metacognitive at all.
The review’s bridge therefore requires construct discipline in both directions.
Neuroscience should not mistake a laboratory metric for the whole of learning. Education should not treat every strategy linked to grades as proof of metacognitive mechanism.
The training literature is uneven
Fleur and colleagues review evidence that direct metacognitive instruction, learning journals and metacognitive prompts can improve aspects of learning and academic performance.
They also note that interventions focused on executive-function training show more mixed outcomes.
This asymmetry leads them to suggest that offline metacognition may be more trainable or more educationally relevant than online processes, although the conclusion remains provisional.
The paper is especially careful around learning journals and prompts.
At the time of the 2021 review, the authors noted that these approaches made up a substantial part of metacognitive training research but that meta-analyses specifically focused on learning journals had not yet been conducted. Individual studies suggested benefits, and the way journals were designed mattered: reflective prompts appeared more useful than simple summaries.
Later work strengthens the prompt side. Lin Guo’s 2022 meta-analysis of experimental studies in computer-based learning environments found positive average effects of metacognitive prompts on self-regulated-learning activities and learning outcomes, with feedback, specificity and adaptability moderating effects.
That later evidence does not retroactively prove every journal intervention effective.
It confirms that design details matter.
A prompt can scaffold a process without creating a trait
Metacognitive interventions are sometimes described as though they permanently increase a general capacity.
The review supports a more cautious interpretation.
A prompt may cause a learner to plan in one environment. A learning journal may create repeated opportunities for reflection. Direct instruction may improve knowledge about strategies. Those changes are valuable without assuming broad transfer.
To claim a general metacognitive trait has changed, researchers would need evidence across tasks, domains and time.
This is one of the places where educational and neuroscience approaches can help each other. Educational studies offer real learning contexts. Laboratory methods offer fine-grained tests of monitoring and transfer.
The bridge becomes strongest when a programme can show both meaningful educational outcomes and changes in the specific metacognitive process it claims to train.
Domain generality remains an open question
Is a person who is metacognitively accurate in one domain also accurate in another?
The review treats this as unresolved.
Some evidence suggests common components. Other findings show domain specificity. Task-based measures themselves can have limited test-retest reliability, as Rahnev’s 2025 methods paper later emphasized.
The question matters enormously for education and organisational assessment.
If metacognition is strongly domain-specific, a single general score will mislead. If meaningful cross-domain components exist, training and assessment might target them more efficiently.
The evidence supports neither an extreme “one trait explains everything” position nor an extreme “nothing transfers” position.
Neuroscience adds mechanism without conferring authority by itself
Brain imaging can identify networks associated with error monitoring, confidence and control.
That does not automatically make a measure more valid.
Neural correlates can help distinguish mechanisms, test models and reveal developmental changes. But the presence of a brain activation does not prove that an educational intervention has improved metacognition in a meaningful way.
Every psychological process has a neural substrate.
The epistemic value comes from whether the neural measurement adds discriminating evidence to the theory.
The review’s cross-disciplinary ambition is strongest when neuroscience and education constrain each other rather than when one is used to decorate the other.
Why European research infrastructure matters here
The paper emerges from an unusually useful institutional intersection.
The University of Amsterdam contributes cognitive and computational research. Amsterdam University of Applied Sciences contributes educational context. The Max Planck Institute for Human Development contributes developmental and decision-science expertise.
That mixture mirrors the methodological bridge the paper argues for.
For the Institute, European authority should not mean attaching a regional source to satisfy geography. It should mean finding research ecosystems that genuinely connect laboratory mechanism, learning environments and human-development questions.
This review does that.
AI makes the old methodological split more urgent
Generative AI can now support planning, explain strategies, provide feedback, summarize errors and prompt reflection.
It can also make performance easier without making the learner more metacognitively capable.
Educational measures may show that students feel more reflective after using an AI coach. Laboratory confidence measures may show no change in calibration. Both findings could be true.
The Institute’s synthesis is that AI-era metacognition research should specify level of measurement before claiming improvement: reported strategy knowledge, observed regulatory behaviour, trial-level monitoring accuracy, academic outcome, transfer or unaided performance.
The 2021 review did not study modern generative AI. Its cross-disciplinary architecture makes the measurement problem legible.
Development changes what a metacognitive measure means
A bridge between educational and neuroscience traditions also has to be developmental.
Children, adolescents and adults do not enter a metacognitive task with the same language, executive resources, experience or social expectations. A young learner may recognize that something is difficult without being able to articulate a strategy. An adolescent may know an effective strategy but not deploy it under motivational or social pressure. An adult expert may regulate efficiently with little conscious verbalization.
That means age differences in self-report, think-aloud richness or executive performance cannot automatically be interpreted as age differences in one underlying metacognitive quantity.
The 2021 review discusses developmental and educational perspectives precisely because mechanism and expression change together. Neuroscience can help trace changes in control and monitoring systems. Education can show how strategy knowledge, feedback and classroom norms shape the same learner over time.
A mature developmental account therefore asks not only “how much metacognition?” but “which component, expressed through which behaviour, at this stage, under which demand?”
Ecological validity is not the opposite of experimental control
Classroom studies are often praised as realistic and laboratory studies criticized as artificial.
That contrast is too crude.
A tightly controlled task can reveal a mechanism that would be impossible to isolate in a classroom. A classroom study can show whether a mechanism survives distraction, motivation, peer dynamics, subject knowledge and real consequences. The two settings answer different questions.
Problems arise when researchers borrow the prestige of one setting to answer a question that belongs to the other. A perceptual confidence task should not be advertised as a complete measure of a student’s learning wisdom. A reflective journal should not be treated as direct access to trial-level monitoring precision.
The strongest programmes can stage evidence across levels: first establish a mechanism under controlled conditions, then test whether an intervention changes behaviour in authentic learning, then examine transfer and persistence. Not every study must do all three. The field needs a portfolio in which the levels constrain one another.
Offline and online are useful labels, not two separate minds
The offline-online distinction can accidentally sound like a claim that humans possess two metacognitive systems, one reflective and one immediate.
The review does not require that conclusion.
An “offline” planning statement can be informed by years of online error experience. An “online” confidence judgment can be biased by a broad self-belief such as “I am bad at mathematics.” Reflection after a task can change the strategy chosen on the next trial. The levels are coupled over time.
For education, this coupling may be exactly where intervention matters. A journal is useful only if reflection later changes attention, strategy choice or help seeking. A confidence signal matters only if it affects control. A classroom norm that welcomes uncertainty can change whether a learner reports an error early enough to act on it.
So the scientific task is not merely to catalogue offline and online measures. It is to study the pathways between them.
Measurement portfolios are better than a single metacognition score
The review points toward a practical principle for institutions: use multiple measures when the decision matters.
A school evaluating a metacognitive programme might combine task performance, delayed learning outcomes, observed strategy use, confidence judgments and carefully designed self-report. A research lab may add computational measures or neural data. A workplace study might include decision quality, error detection and help-seeking behaviour.
Triangulation does not mean that all measures should agree perfectly. If they measure different levels, divergence can be informative. A participant may report improved strategy knowledge without showing better online calibration. That is not automatically a failed intervention; it tells researchers where the change may have stopped.
The danger is creating a composite score so broad that disagreement disappears inside arithmetic. A portfolio should preserve the meaning of each measure and connect each one to a claim.
This is especially important in AI-supported learning, where self-reported ease, assisted output quality and retained competence can move in different directions at the same time.
What the evidence does not show
The review does not establish a single agreed measurement of metacognition.
It does not show that educational self-reports are useless, only that their relationship with outcomes can be limited and construct-dependent.
It does not prove that think-aloud protocols reveal cognition without reactivity.
It does not show that executive-function training reliably improves academic achievement.
Evidence that metacognitive instruction, prompts or journals can help learning does not prove that every intervention works or that effects transfer across subjects and populations.
The paper does not establish whether offline metacognition is intrinsically more trainable than online metacognition; it offers a synthesis consistent with the evidence available at the time.
And neural evidence does not replace behavioural validity.
What changed with this review
The paper makes it harder to say “metacognition was measured” without asking how.
That is a meaningful advance.
A student reflecting in a journal, a participant rating confidence and a brain-imaging study of error signals are not three interchangeable versions of the same measurement.
They are observations made at different levels.
The scientific opportunity is not to force them into one metric.
It is to understand how the levels relate.
A mature field should be able to watch someone learn in a classroom, observe a monitoring process in a controlled task and explain where the two genuinely connect without pretending the bridge has already been completed.
Fleur, Bredeweg and van den Bos show why that bridge is worth building.
Research references
Fleur DS, Bredeweg B, van den Bos W. Metacognition: ideas and insights from neuro- and educational sciences. npj Science of Learning. 2021;6:13. https://doi.org/10.1038/s41539-021-00089-5
Guo L. Using metacognitive prompts to enhance self-regulated learning and learning outcomes: A meta-analysis of experimental studies in computer-based learning environments. Journal of Computer Assisted Learning. 2022;38(3):811–832. https://doi.org/10.1111/jcal.12650
Rahnev D. A comprehensive assessment of current methods for measuring metacognition. Nature Communications. 2025;16:701. https://doi.org/10.1038/s41467-025-56117-0
Fox MC, Ericsson KA, Best R. Do procedures for verbal reporting of thinking have to be reactive? Psychological Bulletin. 2011;137(2):316–344. https://doi.org/10.1037/a0021663