Digital Learning Tools Boost Metacognition on Average, Major Meta-Analysis Finds

Educational technology has long promised to make learners not just better at tasks, but better at thinking about their own thinking. A sweeping new synthesis published in Educational Psychology Review now offers the most comprehensive quantitative answer yet to whether that promise holds. Researchers Yijia Yuan and Yiran Du of the University of Cambridge pooled results from 33 experimental and quasi-experimental studies, encompassing 94 separate effect sizes, to estimate how technology-supported interventions—from simple reflective prompts to generative artificial intelligence—shape learners’ metacognition. Their conclusion is cautiously encouraging: across the literature, technology-supported interventions were associated with moderately higher metacognitive outcomes, with a pooled effect of Hedges’ g = 0.653 and a 95 percent confidence interval running from 0.520 to 0.786. Yet the authors are equally clear about what that number does and does not mean, because the evidence beneath it is far more uneven than the headline statistic suggests.

Metacognition, in the framework the researchers adopted, covers both knowledge about one’s own cognition and the regulation of cognitive activity through planning, monitoring, evaluation, and strategic control. Digital environments are natural arenas for these processes because they flood learners with choices—multiple pathways, abundant resources, automated feedback—and demand that learners judge whether their strategies are working. Technologies can be designed to scaffold regulation directly: prompts can nudge goal setting, dashboards can make progress visible, learning diaries can structure reflection, and feedback systems can highlight the gap between intended and actual performance. With the arrival of intelligent tutoring systems, learning analytics, and now generative AI capable of open-ended explanation and critique, the repertoire of possible support has expanded dramatically. But more capable systems can also perform regulatory work on the learner’s behalf, raising a question that runs through the entire review: does the technology strengthen independent regulation, or quietly absorb it?

To answer such questions rigorously, the team followed PRISMA 2020 reporting standards and searched five major databases—Web of Science, Scopus, ERIC, IEEE Xplore, and PsycINFO—completing searches on 9 July 2026 with no lower publication-year limit. The initial sweep returned 5,326 records. After removing 2,532 duplicates, 2,794 unique records underwent title-and-abstract screening, and 168 reports were assessed in full text. Thirty-three studies ultimately met all eligibility criteria: each had to involve students in formal K–12 or post-secondary settings, compare a technology-supported intervention against a control or business-as-usual condition, and report a quantitative metacognitive outcome from which a standardised effect size could be derived. Two researchers screened and coded everything independently, achieving Cohen’s kappa values between 0.84 and 0.94 on categorical variables and intraclass correlations above 0.94 on continuous quantities—agreement figures that signal unusually disciplined extraction.

The statistical engine of the study was a three-level random-effects meta-analysis, a model chosen because the 33 studies contributed 94 effect sizes and multiple effects from the same study cannot be treated as independent. Level 1 of the model captured sampling variance, Level 2 captured heterogeneity among effect sizes within studies, and Level 3 captured heterogeneity between studies. The pooled estimate of g = 0.653 was highly significant, but the heterogeneity was substantial: total I² reached 72.7 percent, and virtually all of the excess variance sat at the between-study level, with the within-study component estimated at effectively zero. A 95 percent prediction interval stretching from roughly −0.02 to 1.32 drives the point home. In a comparable future implementation, the true effect could plausibly be negligible—or large. The average, in other words, describes a scattered landscape rather than a uniform intervention effect.

Perhaps the most revealing part of the analysis is a descriptive map of what the interventions actually did. Evaluation and reflection were explicitly supported in 29 of 33 studies, or 87.9 percent; monitoring and strategic control each appeared in 26 studies, or 78.8 percent; but planning was represented in only 19 studies, or 57.6 percent. The pattern suggests that most digital interventions intervene after learning is under way—prompting learners to inspect progress, self-assess, and revise—while far fewer help them set goals and select strategies before starting. Because goals and task analysis provide the foundation on which monitoring and evaluation depend, the authors argue that future designs should attend to the full regulatory cycle rather than its back half.

The outcome evidence was even more lopsided. Metacognitive regulation accounted for 77 of the 94 effect sizes, or 81.9 percent, while metacognitive knowledge contributed just 11 effects and composite measures six. More striking still, 90 of 94 effects—95.7 percent—rested on self-report instruments. Performance-based, judgement-based, or calibration indicators produced only three effects, trace or process data contributed one, and not a single study used researcher-rated task products. This means the pooled estimate primarily summarises changes in what learners say about their own planning, monitoring, and reflection, not independently demonstrated regulatory skill. Self-reports are legitimate data, and learners’ perceptions of their strategy use can genuinely change in response to instruction, but perceived regulation is not equivalent to accurate or effective regulation enacted during performance—a distinction long emphasised in the metacognition literature.

The exploratory moderator analyses generated the study’s most newsworthy, and most qualified, findings. Three implementation features cleared the conventional unadjusted significance threshold: interventions with explicit technology-use guidance outperformed unguided ones (g = 0.760 versus 0.479), immediate feedback beat delayed feedback (g = 0.744 versus 0.408), and system-directed learner control beat learner-directed control (g = 0.780 versus 0.499). But none of these survived Benjamini–Hochberg false-discovery-rate correction across the eight moderator tests, with each adjusted p value landing at 0.095. Crucially, the three features overlapped heavily in practice—14 of the 20 guided studies also provided immediate feedback—so their apparent effects may reflect a bundled instructional style rather than three separable design principles. The authors explicitly frame these patterns as hypothesis-generating rather than as established moderator effects.

Technology type, notably, did not significantly moderate outcomes. Generative-AI interventions showed the largest descriptive subgroup estimate at g = 0.831, compared with 0.643 for non-AI technologies and 0.346 for non-generative AI, but those subgroups contained only nine and four studies respectively, drawn from heterogeneous systems. The findings therefore do not support any claim that AI, or generative AI in particular, is intrinsically superior for building metacognition. Educational level, subject area, and intervention duration also failed to explain meaningful variation. The robustness of the headline result, by contrast, was impressive: cluster-robust inference reproduced the same point estimate, imposing assumed within-study correlations from 0.30 to 0.90 shifted the pooled estimate only between 0.640 and 0.651, leave-one-study-out estimates ranged from 0.625 to 0.675, and excluding the two studies rated at serious risk of bias left the association intact at g = 0.624.

The risk-of-bias picture tempers enthusiasm further. Among the 33 studies, only nine used randomised controlled designs; 22 quasi-experimental studies were judged at moderate risk under ROBINS-I, two at serious risk, and no study at all received a low overall risk rating. A cluster-robust Egger-type precision test found no statistically detectable evidence of small-study effects, though the authors interpret this as an absence of evidence rather than evidence that publication bias is absent. These caveats, combined with the self-report dominance, mean the pooled effect is best read as a positive average association in reported metacognitive regulation rather than proof of uniformly improved metacognitive capability.

The practical upshot is a design philosophy rather than a shopping list. The authors argue that technology selection should begin with the regulatory problem to be solved—goal setting, monitoring, evaluation, or strategy adaptation—rather than with the technological label, and that AI systems in particular should be judged by which regulatory decisions remain with the learner and how support is faded over time. For researchers, the priorities are larger controlled studies, transparent descriptions of what each design feature is meant to elicit, and multimethod assessment combining self-reports with calibration judgements, performance indicators, and behavioural traces. As generative AI spreads through classrooms, the difference between a tool that cultivates independent thinkers and one that induces what other researchers have called metacognitive laziness may depend on exactly those choices. This synthesis gives educators a genuine, if qualified, reason for optimism—and a precise agenda for finding out what actually works.

Subject of Research: Effects of technology-supported educational interventions, including AI tools, on learners' metacognition

Article Title: Technology-Supported Interventions and Learners’ Metacognition: A Three-Level Meta-Analysis

Article References: Yuan, Y., & Du, Y. (2026). Technology-Supported Interventions and Learners’ Metacognition: A Three-Level Meta-Analysis. Educational Psychology Review, 38(1), Article 130. https://doi.org/10.1007/s10648-026-10227-3

Image Credits: AI Generated

DOI: 10.1007/s10648-026-10227-3

Keywords: metacognition, educational technology, meta-analysis, self-regulated learning, generative AI, artificial intelligence in education, learning analytics, feedback, scaffolding, self-report measures, educational psychology, learner control