Truace
Subscribe
70/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Meta-analysis of 21 studies finds AI gives blended-learning students a medium boost

Study title: Unlocking the potential of artificial intelligence in improving learning achievement in blended learning: a meta-analysis

Evidence type
Systematic review or meta-analysis
Sample
n = 2,873
Sources
1 sourcePeer-reviewed
Published
3 min read
Truvace record
TRV-2026-2830cite or check it

Study title: Unlocking the potential of artificial intelligence in improving learning achievement in blended learning: a meta-analysis

Researchers pooled 21 studies covering 2,873 students to test whether adding AI to blended learning (a mix of online and in-person teaching) improves learning achievement. AI-enhanced courses showed a medium effect (Hedges' g = 0.50, 95% CI 0.27 to 0.74) compared with control groups without AI. Sixteen studies used true experimental designs and five were quasi-experimental.

Read the study abstract

Introduction While several studies investigated the effect of blended learning on students’ learning achievement, scant information exists on whether using artificial intelligence (AI) in blended learning could further contribute to the obtained effect. Methods To effectively address the challenges and opportunities presented by blended learning and AI, the present study conducts a meta-analysis to systematically examine the impact of AI-enhanced blended learning on students’ learning achievement, considering th…

62Gain side

How the two sides scoreClose call

61Problem side

The gain

A meta-analysis of 21 studies with 2,873 students found AI-enhanced blended learning raised learning achievement by a medium margin (Hedges' g = 0.50), with personalized AI systems reaching g = 0.88.

The problem

The same meta-analysis of 21 blended-learning studies found AI's benefit faded to g = 0.08 in interventions lasting a semester to a year, and the one early-childhood study showed a significant negative effect (g = -0.31).

A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.

What the scores mean

G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.

P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.

In brief

The benefit depended heavily on context. Personalized AI systems (four studies) had a large effect (g = 0.88), intelligent tutoring systems (16 studies) a small-to-medium one (g = 0.42), and the single chatbot study a non-significant one (g = 0.34). Higher education (nine studies) showed g = 0.7, while the one early-childhood study found a significant negative effect.

The results vary widely between studies (I² about 89%), and the prediction interval runs from -0.57 to 1.57, so a future study could find no benefit. Effects were largest in short interventions (g = 1.14 over one week to one month) and nearly vanished over a semester to a year (g = 0.08). Tests also pointed to publication bias, with a bias-adjusted estimate of g = 0.32.

Main points

  1. The pooled effect across 21 studies and 2,873 participants was g = 0.50, with a 95% confidence interval of 0.27 to 0.74.
  2. Personalized AI systems (k = 4) had g = 0.88; intelligent tutoring systems (k = 16) had g = 0.42; one chatbot study had a non-significant g = 0.34.
  3. Higher education (k = 9) showed g = 0.7 and secondary education (k = 9) g = 0.40; primary education (k = 2) was medium but non-significant.
  4. By subject, teacher training (one study) reached g = 0.99 and ICT and engineering (five studies) g = 0.88, while mathematics showed only g = 0.11.
  5. Interventions of one week to one month had g = 1.14, one month to a semester g = 0.53, and a semester to a year g = 0.08.
  6. Studies with 250 or fewer participants had g = 0.57; larger studies had a non-significant g = 0.24.
  7. Rank-correlation and regression tests suggested publication bias; after removing six outliers and trim-and-fill, the estimate was g = 0.3181.

What the review found when it combined the studies

These are the review’s own pooled results, copied from its abstract. Truvace did not recalculate them.

  • Student learning achievement with intelligent tutoring systems in blended learning

    Hedges' g 0.42 (95% CI 0.15 to 0.69)

    16 studies · search to June 01, 2024

    The review’s own words
    Intelligent Tutoring Systems (ITS) ( k = 16) studies reported a small-to-medium and statistically significant effect size (Hedges’ g = 0.42, 95% CI [0.15, 0.69], p = 0.002)

  • Student learning achievement with personalized AI systems in blended learning

    Hedges' g 0.88 (95% CI 0.48 to 1.27)

    4 studies · search to June 01, 2024

    The review’s own words
    studies that used a personalized AI system ( k = 4) reported a large effect size (Hedges’ g = 0.88), which was statistically significant ( p = 0.001, 95% CI [0.48, 1.27])

  • Student learning achievement with AI in blended learning after outlier removal and trim-and-fill adjustment

    Hedges' g 0.3181 (95% CI 0.1672 to 0.4690)

    18 studies · search to June 01, 2024

    The review’s own words
    the Trim-and-Fill adjusted estimate increased to g = 0.3181, with a narrower confidence interval [0.1672, 0.4690]

  • Student learning achievement with AI in blended learning compared with no AI

    Hedges' g 0.50 (95% CI 0.27 to 0.74)

    21 studies · 2,873 people · search to June 01, 2024

    The review’s own words
    The overall pooled effect size of the 21 studies shows a medium effect size (Hedges’ g = 0.50), with a 95% confidence interval of [0.27, 0.74]

Code checks every number above against the review’s text before it is shown. The short labels are an AI model’s reading of that text, so the review’s own sentence sits under each one. A pooled figure is only as strong as the studies behind it. How Truvace uses pooled results

The full story

The team searched Science Direct, IEEE Xplore, Taylor & Francis, Scopus and Web of Science for peer-reviewed English-language studies published from 2011, with the last search on June 1, 2024. Of 538 records, 307 were duplicates, 115 full texts were assessed and 21 met the criteria, which required an AI intervention in blended learning and a control condition. Effect sizes were computed with Hedges' g in Comprehensive Meta-Analysis V.4.

Subgroup differences were marked. Meta-regression showed early childhood had a significantly negative coefficient, mathematics a significant negative coefficient of -0.38, and semester-to-year interventions a coefficient of -1.28. The authors suggest this may reflect a novelty effect or weaker implementation over long periods, though this is interpretation rather than a tested cause.

Publication-bias checks gave mixed signals. Kendall's tau (0.4571, p = 0.0033) and a regression test indicated small studies reporting larger effects were overrepresented, yet a p-curve analysis indicated evidential value. The authors call for longitudinal work and for evidence from primary schools, where data are thin.

What this doesn’t fix

Heterogeneity was very high (I² about 89%) with a prediction interval of -0.57 to 1.57, several subgroups rested on one or two studies, and tests suggested publication bias; only English-language journal papers were searched.

Sources

1 on file
  1. Frontiers in Psychologydoi:10.3389/fpsyg.2025.1691414
    Peer-reviewed2025-10-31

Talk it over

Four questions to start onThis record found a gain and a problem in the same place. Which matters more to you, and who might answer differently?

Start with a question

Meta-analysis of 21 studies finds AI gives blended-learning students a medium boosthttps://truvace.com/item/unlocking-the-potential-of-artificial-intelligence-in-improving-learning-achieve

  1. This record found a gain and a problem in the same place. Which matters more to you, and who might answer differently?

    The gain A meta-analysis of 21 studies with 2,873 students found AI-enhanced blended learning raised learning achievement by a medium margin (Hedges' g = 0.50), with personalized AI systems reaching g = 0.88.

    The problem The same meta-analysis of 21 blended-learning studies found AI's benefit faded to g = 0.08 in interventions lasting a semester to a year, and the one early-childhood study showed a significant negative effect (g = -0.31).

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on a systematic review or meta-analysis with a sample of 2,873, from Frontiers in Psychology.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    The record’s own caveat Heterogeneity was very high (I² about 89%) with a prediction interval of -0.57 to 1.57, several subgroups rested on one or two studies, and tests suggested publication bias; only English-language journal papers were searched.

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Science