A 2026 meta-analysis of 35 experimental and quasi-experimental studies found that AI-supported instruction raised STEM achievement test results by an average of 0.67 standard deviations (Hedges' g, 95 percent confidence interval 0.49 to 0.85) compared with traditional teaching. The gains were uneven: largest in high school, smallest in middle school, and based on test scores rather than on whether students became more interested in STEM. In Canada, the federal CanCode program is funding a fifth phase with CAD 30 million for coding and AI skills in kindergarten to grade 12.
Key takeaways
- The meta-analysis, published June 15, 2026 in the Journal of Intelligence, pooled 35 studies from 2005 to 2025 and reports a moderate-to-large average effect on achievement (g = 0.67).
- By school level, the effect was 1.10 in high school, 0.58 at university, 0.47 in elementary school, and 0.39 in middle school.
- Differences between science, mathematics, and technology or engineering were not statistically significant (p = 0.088).
- The studies measured test performance, so the result says little about inspiring students or about hands-on AI activities.
- CanCode phase five targets 1.1 million students and 76,000 teachers over two fiscal years.
What the meta-analysis pooled
Doğan, Kılıç, Kalınkara, and Talan searched Web of Science, Scopus, ERIC, ScienceDirect, and a scholarly search engine for studies published between 2005 and 2025. They kept experimental and quasi-experimental studies in which an AI-supported instructional intervention in a STEM subject was compared with traditional teaching. The outcome was academic achievement, mostly scores on tests and examinations. One effect size was taken from each of the 35 studies, and a random-effects model was used because the studies differed. The open-access paper reports high heterogeneity, with a fixed-effect heterogeneity statistic of about 83 percent, which indicates that much of the variation between studies goes beyond chance.
"AI-supported" covers a range of tools, from tutoring and adaptive systems to generative AI, and the pooled number blends them. The authors call for more work on generative AI in particular.
Results by level, subject, and duration
| Grouping | Category | Effect size (g) | Note |
|---|---|---|---|
| School level | High school | 1.10 | Largest effect |
| School level | University | 0.58 | Moderate |
| School level | Elementary school | 0.47 | Lower |
| School level | Middle school | 0.39 | Smallest effect |
| Subject | Science | 0.68 | Differences between subjects not significant |
| Subject | Mathematics | 0.65 | Not significant |
| Subject | Technology and engineering | 0.50 | Not significant |
| Duration | More than one month, up to two months | 0.83 | Largest effect |
| Duration | Five hours or less | 0.62 | Significant |
| Duration | More than one day, up to seven days | 0.26 | Not significant |
The authors note that longer was not always better, so the pattern does not support a simple rule such as "run it for a month." The high school figure is nearly three times the middle school one. The authors link this to learner readiness, an explanation they propose but did not test directly.
What the result does not say
Three limits matter for a teacher deciding on activities. First, the outcome is test achievement. The review does not measure interest, identity, or the intention to study STEM later, so it cannot show that an activity inspires students. Second, most included interventions were short or medium in length, and the authors say long-term persistence is unknown. Third, the Begg test for publication bias came out at p = 0.0517, just above the usual cutoff, so the authors advise caution even though the other checks found little sign of bias.
The review also does not test unplugged or build-your-own-model activities. A class that trains a simple classifier on its own photos is learning about AI, which differs from a class where AI tutors learn about fractions. The activities in the related article on classroom AI activities are of the first kind, and the framework behind them presents them as untested scenarios.
Canada's CanCode: public money for coding and AI skills
On July 8, 2026, the Minister of Industry announced CAD 30 million for the fifth phase of CanCode, the Government of Canada's program for digital skills in kindergarten to grade 12. The stated aim is learning opportunities for 1.1 million students and training for 76,000 teachers. The government news release describes coding and foundational AI knowledge, and the applicant guide lists coding, AI, data analytics, and digital content creation.
Funding goes to not-for-profit organizations, not to schools directly, and activities must be free to participants. Applicants planning to teach responsible AI use receive extra consideration, and girls, Black youth, Indigenous youth, youth with disabilities, and youth in rural, remote, or northern communities are named priority groups. The call closed on August 6, 2026, with projects running from fall 2026 to March 31, 2028. The announcement sets reach targets; it does not report learning results.
Questions to ask before adding an AI activity to a STEM unit
- What will students be able to do without the tool afterward, and how will that be checked?
- Is the goal learning content with AI support, or learning how AI works?
- Does the plan fit the age group, given that effects in the review varied by level?
- Who sees student data, and does district or provincial policy allow the tool?
- Is there a comparison, such as a parallel class or a before-and-after quiz, to judge the result?
FAQ
Does AI make students better at science and math?
On average, studies in this meta-analysis found higher test results with AI-supported teaching than with traditional teaching. The spread between studies was large, so a given tool in a given class may do better or worse.
Does the review show that AI inspires students to choose STEM?
No. It looked at achievement scores. Interest and career intentions were outside its scope, and the authors recommend studying motivation and other non-test outcomes in future work.
Which grade levels benefit most?
High school students showed the largest average effect and middle school students the smallest. Studies at each level used different tools and designs, so the comparison is suggestive, not a ranking of grades.
Can schools apply to CanCode directly?
The program funds not-for-profit organizations that deliver training. The 2026 to 2028 call closed on August 6, 2026, so schools interested in its programming would look for approved organizations active in their area.
Related reading
- Classroom AI activities from an international framework
- The wider impact of AI in education
- AI and academic writing skills
- Foundations of elementary education
More in AI in education.
Last verified: October 2026
Sources
- The Impact of Artificial Intelligence-Supported Instruction on Student Learning in STEM: A Systematic Review and Meta-Analysis, Doğan Y, Kılıç Z, Kalınkara Y, Talan T, Journal of Intelligence 14(6):109, 2026
- Government of Canada invests $30 million to launch fifth phase of CanCode program, Innovation, Science and Economic Development Canada, July 8, 2026
- CanCode program 2026-2028: Applicant guide, Innovation, Science and Economic Development Canada, 2026