Does generative AI improve self-regulated learning?
A meta-analysis of 17 comparative studies found a positive average effect, but results varied sharply: heterogeneity reached 94.85% and the prediction interval included harm as well as benefit. Almost all evidence came from higher education.
The short version
3 key points
- 1.Seventeen comparative studies listed 1,359 participants; 16 studies were in higher education and only one was in primary education.
- 2.The pooled effect was Hedges' g = 0.745, but heterogeneity was 94.85% and the 95% prediction interval ranged from -1.03 to 2.51, so a new setting could plausibly show no benefit.
- 3.Removing two unusually large effects reduced the pooled estimate to 0.49. The review did not establish that students retained self-regulation after AI support was removed.
The answer: a positive average, not a reliable promise
Students receiving generative-AI-supported teaching showed better self-regulated-learning scores on average than comparison groups in a new meta-analysis. Across 17 studies, the pooled standardized effect was Hedges' g = 0.745, with a 95% confidence interval from 0.315 to 1.174. That is a moderately large average association and remained positive when any single study was removed. It supports careful trials of AI as a scaffold for planning, monitoring, reflection and learning strategy.
It does not show that an AI assistant will improve autonomy in every classroom. Statistical heterogeneity was 94.85%, meaning most observed variation was not ordinary sampling noise. The 95% prediction interval ran from -1.03 to 2.51 and crossed zero by a wide margin. In practical terms, the average is compatible with settings that benefit substantially and settings that show no benefit or worse outcomes. Teachers should therefore evaluate the design and local result rather than importing the pooled number as an expected effect.[1]
What the meta-analysis included
Researchers at Universiti Kebangsaan Malaysia searched Scopus and Web of Science for English-language peer-reviewed studies published from January 2023 through 30 June 2026. The search, conducted in July 2026, returned 649 records. Eligible studies had to compare a generative-AI intervention with a control condition, measure self-regulated learning or one of its direct dimensions, and report enough statistics to calculate an effect size. Two reviewers screened and coded studies independently, consulting a third when necessary.
Seventeen studies met the criteria. Their appendix lists 1,359 participants in total, with individual study samples ranging from 31 to 192. Sixteen studies were conducted in higher education and one in primary education. The interventions included conversational and adaptive AI, often ChatGPT, across language and STEM subjects. The team used the Mixed Methods Appraisal Tool and found recurring problems: convenience or single-site samples, incomplete reporting of randomisation and allocation, little assessor blinding, residual confounding and some incomplete outcome data.[1]
How the effect was calculated
The authors calculated Hedges' g from post-intervention group means where possible, using adjusted estimates, change scores or convertible test statistics when needed. Positive values represented better self-regulation in the AI-supported group. When a study reported several relevant outcomes but no overall score, the team combined them into a study-level estimate and assumed a within-study correlation of 0.50. Repeating the analysis with correlations of 0.20 and 0.80 barely changed the pooled result.
A random-effects model allowed the underlying effect to differ among studies. The authors used restricted maximum likelihood for between-study variance and a Hartung-Knapp adjustment because only 17 independent studies were available. They also tested publication-bias signals, removed each study in turn, inspected influential cases and repeated the analysis without effects above 2.0. This is a more cautious synthesis than simply averaging published scores, but no statistical method can compensate for a small, heterogeneous and mostly university-based evidence base.[1]
The large headline effect became smaller under stress testing
Two studies reported effects above Hedges' g = 2.0. Excluding them reduced the pooled estimate from 0.74 to 0.49, with a 95% confidence interval from 0.24 to 0.74. Heterogeneity also fell, but remained high at 83.27%. Removing the single most influential study produced g = 0.60, still positive, while heterogeneity remained above 90%. The direction of the average is therefore reasonably stable; its magnitude and portability are not.
Funnel-plot inspection and Egger's test did not identify statistically significant asymmetry, but 17 studies give limited power to detect publication bias. The review also searched only two databases and excluded non-English work, ERIC and PsycINFO. The apparent stability of the pooled estimate should not be confused with complete coverage of the field, especially when rapid educational experimentation may appear first in conference proceedings, institutional reports or languages outside the search.[1]
Planning, motivation and behaviour all looked positive—but not different
Fourteen studies contributed 34 effects to a dimension-level model. Point estimates were positive for cognitive regulation, motivational regulation, metacognitive regulation and behavioural regulation. Yet a robust comparison found no statistically reliable difference among those four dimensions. The motivational estimate was based on only three studies, so the absence of a difference is not evidence that every type of self-regulation benefits equally.
The analysis also reported larger effects in randomised trials than quasi-experimental studies and in interventions lasting five to eight weeks than longer interventions. Both findings need restraint. Heterogeneity within the randomised subgroup was 96.92%, duration subgroups contained few studies, and the duration finding survived only nine of 17 leave-one-out tests. After adjusting for duration, the research-design moderator was no longer statistically significant. These patterns are hypotheses for better trials, not instructions to choose a particular course length.[1]
What teachers and students should do with the result
A useful classroom design would make the AI prompt students to set a goal, explain a strategy, compare alternatives, check evidence and reflect on errors. It would preserve teacher feedback and peer discussion, disclose when the model is uncertain and ask students to complete some planning and monitoring without assistance. Schools and universities should measure whether students can repeat the process independently, rather than treating faster completion or a positive attitude survey as proof of stronger self-regulation.
Students should understand the difference between external support and an internalised skill. If the model plans, checks and revises every task, the immediate product may improve while the learner practises less regulation. The review itself makes that distinction: current evidence mostly measures performance while AI is present. It does not establish durable transfer after support is withdrawn, and it offers almost no basis for claims about primary or secondary school pupils.[1]
What would change the assessment
Confidence would rise with preregistered multicentre trials that use representative students, report allocation and attrition clearly, and test a defined instructional design rather than access to a general chatbot. Outcomes should include observed planning and monitoring behaviour, validated self-regulation scales, course attainment, equity, inappropriate reliance and delayed tests after AI support is reduced or removed. Researchers should publish null and harmful results as well as successes.
The authors reported no external funding or conflicts and disclosed using ChatGPT only for language editing. For now, their synthesis supports a measured conclusion: generative AI can scaffold self-regulated learning in some settings, but the average effect hides extremely variable results. Teachers should use the tool to make students' reasoning more visible and independent—not to take over the regulatory work the lesson is meant to develop.[1]
What this means for people
- Teachers can use AI to prompt planning and reflection, but should test whether students can perform those processes after support is withdrawn.
- Students may complete work more effectively with an AI scaffold without developing durable independence, so assisted performance and learned skill must be measured separately.
- School and university leaders should not apply a higher-education-heavy average to younger pupils without age-appropriate trials.
Global context
The review team was based in Malaysia, while the included studies spanned varied subjects and settings. Sixteen of 17 studies were in higher education, the search was English-only and country-level transfer was not established, so local pedagogy, language and access conditions matter.
What the evidence does not yet show
- Only 17 studies and 1,359 listed participants, with small moderator subgroups
- Very high heterogeneity and a prediction interval that crossed zero
- Sixteen studies in higher education and only one in primary education
- Search limited to Scopus, Web of Science and English-language peer-reviewed work
- No pooled evidence that self-regulation persisted after AI support was withdrawn
What to watch next
- Preregistered multicentre trials with delayed post-tests
- Scaffold-fading designs that measure independent performance
- Evidence in primary and secondary schools
- Validated behavioural as well as self-report outcomes
- Publication of null, harmful and equity-disaggregated results
Living evidence record
Impact record IAI-19CZQ68
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
15 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Information published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 15 October 2026
Research question, topics and regions
Whether comparative studies show that generative-AI-supported learning improves students' planning, monitoring, motivation and learning behaviour
Topics: Generative AI · Self-regulated learning · Higher education · Meta-analysis · Teaching · Student autonomy
Regions: Malaysia · Asia · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Watch the background, then check the evidence
UNESCO
AI-generated illustration · not a video frame
YouTube loads only when you choose. Press play in the player to start. If playback is unavailable, use “Watch on YouTube” below.
Education & human agency · Background explainer
Published 2025-01-23 · 8 min 30 sec
Teacher to Teacher: Is AI reshaping education for the better?
A conversation about how teachers can guide AI use while preserving pupils’ judgement. Watch for the difference between helping a learner think and supplying an answer.
Evidence check: Teachers’ perspectives provide context. They do not establish that a particular AI tool improves learning outcomes.
Source checked 14 October 2026. Check available captions on the publisher’s video page.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Education
What do 394 higher-education studies show about generative AI?
A peer-reviewed scoping review finds that universities mostly used generative AI for feedback, tutoring and writing. The evidence is substantial but still dominated by surveys and short-term measures, with little attention to equity or learning processes.
10 min · 1 source
Education
Does AI help musicians practise independently?
A review of 16 heterogeneous studies maps how recording tools, error detectors and generative assistants may support practice. Short-term findings are encouraging, but no study tested whether gains persist after the support is removed.
6 min · 1 source
Education
Did ChatGPT improve writing scores in a 16-week Taiwan course?
Twelve university students used ChatGPT throughout a five-stage inquiry-writing course. Their closed-book writing scores rose by 2.79 points, but the change was not statistically significant and the study had no comparison group.
8 min · 1 source
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.