Truace
Subscribe
70/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Meta-analysis of 35 studies finds AI-supported STEM teaching lifts student test results

Study title: The Impact of Artificial Intelligence-Supported Instruction on Student Learning in STEM: A Systematic Review and Meta-Analysis

Evidence type
Systematic review or meta-analysis
Sample
n = 35
Sources
1 sourcePeer-reviewed
Published
2 min read
Truvace record
TRV-2026-2782cite or check it

Study title: The Impact of Artificial Intelligence-Supported Instruction on Student Learning in STEM: A Systematic Review and Meta-Analysis

A systematic review and meta-analysis published in the Journal of Intelligence in June 2026 pooled 35 experimental and quasi-experimental studies of AI-supported instruction in STEM education. Under a random-effects model the overall effect on student achievement was g = 0.67 (95% CI 0.49 to 0.85), a moderate to large improvement over traditional teaching.

Read the study abstract

The rapid integration of artificial intelligence (AI) technologies into educational contexts has introduced innovative instructional approaches, particularly within Science, Technology, Engineering, and Mathematics (STEM) education. Although an increasing number of empirical studies have examined AI-supported instruction, existing findings remain heterogeneous, making it difficult to draw firm conclusions about its overall effectiveness. This study aims to systematically synthesize experimental and quasi-experim…

The gain

A meta-analysis of 35 experimental and quasi-experimental studies found students taught with AI-supported instruction in STEM scored higher than peers taught traditionally (Hedges' g = 0.67, 95% CI 0.49 to 0.85), with the largest gain among high schoolers (g = 1.099).

In brief

The effect varied by school level. High school students gained most (g = 1.099), while the effect was smaller at elementary school (g = 0.465), middle school (g = 0.392) and university (g = 0.578). Interventions lasting more than one month and up to two months did best (g = 0.833), whereas those lasting one to seven days showed a weaker effect that was not statistically significant (g = 0.256).

Results were highly uneven between studies, with heterogeneity of about 83% in the fixed-effect analysis. Differences between science (g = 0.676), mathematics (g = 0.650) and technology/engineering (g = 0.501) were not statistically significant, so the authors say they should be read with caution.

Main points

  1. Searches of Web of Science, Scopus, ERIC, ScienceDirect and Google Scholar covering 2005 to 2025 turned up 450 records; 35 studies met the inclusion criteria.
  2. Random-effects estimate: g = 0.670 (95% CI 0.491 to 0.848); fixed-effect estimate: g = 0.593.
  3. High school students gained most (g = 1.099), versus g = 0.465 in elementary school, 0.392 in middle school and 0.578 at university.
  4. Interventions lasting over one month and up to two months showed the largest effect (g = 0.833); five-hour-or-shorter programmes also had a significant effect (g = 0.621).
  5. Interventions lasting one to seven days showed a weaker effect that was not significant (g = 0.256, p = 0.070).
  6. Science (g = 0.676) and mathematics (g = 0.650) scored slightly above technology and engineering (g = 0.501), but the gap was not statistically significant (p = 0.088).
  7. Publication-bias checks (Egger p = 0.095, Begg p = 0.0517, trim-and-fill with one imputed study) suggested little distortion; the fail-safe N was 2,404.

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 outcomes in STEM education with AI-supported instruction compared with control instruction

    Hedges' g 0.67 (95% CI 0.49 to 0.85)

    35 studies · search to 2025

    The review’s own words
    statistically significant overall positive effect of AI-supported instruction on student learning outcomes in STEM education (g = 0.67, 95% CI [0.49, 0.85], p < 0.001)

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 authors searched English and Turkish sources in September 2025 and screened 364 records after removing 86 duplicates. They retrieved 43 full-text reports, excluded 8 for insufficient data or poor methodological quality, and analysed 35 studies, 29 of them published after 2020. Three researchers coded the studies, with inter-rater agreement of κ = 0.93.

Only one effect size per study was used, taken from the outcome closest to overall academic achievement, calculated from post-test means and standard deviations. The AI tools covered included intelligent tutoring systems, adaptive learning platforms, chatbots and simulations. Class size also mattered: effects were largest in medium classes of 50 to 99 students (g = 0.687) and smaller in classes of 49 or fewer (g = 0.448).

The authors read the results as conditional rather than universal: AI worked best within well-designed instruction, with teacher involvement and matched to learners' readiness. They call for studies of long-term retention, implementation quality and generative AI.

Sources

1 on file
  1. Journal of Intelligencedoi:10.3390/jintelligence14060109
    Peer-reviewed2026-06-15

Talk it over

Four questions to start onWho gets this benefit, and who was left out of the study or story?

Start with a question

Meta-analysis of 35 studies finds AI-supported STEM teaching lifts student test resultshttps://truvace.com/item/the-impact-of-artificial-intelligence-supported-instruction-on-student-learning-

  1. Who gets this benefit, and who was left out of the study or story?

    The gain A meta-analysis of 35 experimental and quasi-experimental studies found students taught with AI-supported instruction in STEM scored higher than peers taught traditionally (Hedges' g = 0.67, 95% CI 0.49 to 0.85), with the largest gain among high schoolers (g = 1.099).

    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 35, from Journal of Intelligence.

    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 Studies varied widely in their results (I² about 83%), much of the variance is unexplained, and most interventions were short to medium term, so long-term effects on learning are unknown.

    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 Education