aceModerate evidence
More
AI tutors that also support emotions boosted students' knowledge more, a review of 54 studies finds
Study title: Emotional Artificial Intelligence in Education: A Systematic Review and Meta-Analysis
- Evidence type
- Systematic review or meta-analysis
- Sample
- n = 6,151
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-2828cite or check it
Study title: Emotional Artificial Intelligence in Education: A Systematic Review and Meta-Analysis
A systematic review and meta-analysis in Educational Psychology Review screened 13,879 records and covered 172 empirical articles from 2000 to 2025. Of these, 54 controlled or quasi-experimental studies in 49 articles were pooled. AI that combined learning help with emotional support produced a large knowledge gain (g = 0.88), against a medium gain (g = 0.52) for AI offering only learning help.
Read the study abstract
Abstract Research on emotional artificial intelligence (emotional AI) in education—AI that detects learners’ emotions and/or provides emotional support and AI that produces affective learning outcomes but does not directly provide emotional support—has grown rapidly, yet quantitative syntheses are scarce. We meta-analyzed empirical studies reporting outcomes in knowledge, skills, affect, perception, and behavior. A meta‑analysis was conducted based on 172 articles on empirical research published between 2000 and…
The gain
A meta-analysis of 54 controlled studies found AI that paired cognitive help with emotional support raised students' knowledge scores by g = 0.88, versus g = 0.52 for AI offering cognitive help alone.
In brief
The pooled studies covered 1,684 participants in the integrated-support group of studies and 4,400 in the cognitive-only group. Cognitive-only AI also improved skills (g = 0.44), affect (g = 0.48) and perception (g = 0.85), while integrated support did less well on perception (g = 0.18, confidence interval crossing zero).
Evidence on emotional support alone is thin: only two studies qualified and they were not pooled. For affect, a test showed funnel-plot asymmetry, and a fixed-effects correction cut the pooled effect from 0.48 to 0.20, so some results may be overstated.
Main points
- The meta-analysis pooled 54 experimental studies from 49 articles, drawn from 172 empirical articles published between 2000 and 2025.
- Integrated cognitive and emotional AI support gave g = 0.88 on knowledge acquisition (95% CI 0.36 to 1.40, 10 studies); cognitive-only support gave g = 0.52 (12 studies).
- Cognitive-only AI raised skills by g = 0.44, affect by g = 0.48 across 31 studies, and perception by g = 0.85.
- Integrated support showed only g = 0.18 on perception, with a confidence interval from -0.41 to 0.77.
- Individual studies sometimes found harm: a chatbot with empathic messages lowered primary pupils' maths confidence (g = -0.17), and students felt more anxious with a pedagogical agent than with a human teacher (g = -0.28).
- Most studies were in tertiary education (66.3% of 172) and humanities and social science (52.9%); about 39% had fewer than 50 participants.
- Only two studies tested emotional-only support, so no pooled effect was calculated.
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.
Knowledge acquisition with cognitive-only AI support compared with control
Hedges' g 0.52 (95% CI 0.17 to 0.87)
12 studies
The review’s own words
cognitive-only support yielded a medium effect on knowledge acquisition ( g = 0.52, 95% CI [0.17, 0.87])
Skill acquisition with cognitive-only AI support compared with control
Hedges' g 0.44 (95% CI 0.19 to 0.68)
11 studies
The review’s own words
a near-medium effect on skill ( g = 0.44, 95% CI [0.19, 0.68])
Affective experience with cognitive-only AI support compared with control
Hedges' g 0.48 (95% CI 0.33 to 0.63)
31 studies
The review’s own words
affect ( g = 0.48, 95% CI [ 0.33, 0.63])
Learner perception with integrated cognitive and emotional AI support compared with control
Hedges' g 0.18 (95% CI -0.41 to 0.77)
The review’s own words
a small effect on perception ( g = 0.18, 95% CI [-0.41, 0.77])
Knowledge acquisition with integrated cognitive and emotional AI support compared with control
Hedges' g 0.88 (95% CI 0.36 to 1.40)
10 studies
The review’s own words
Integrated cognitive and emotional support showed a large effect on knowledge acquisition ( g = 0.88, 95% CI [0.36, 1.40])
Learner perception with cognitive-only AI support compared with control
Hedges' g 0.85 (95% CI 0.50 to 1.20)
10 studies
The review’s own words
a large effect on perception ( g = 0.85, 95% CI [0.50, 1.20])
Skill acquisition with integrated cognitive and emotional AI support compared with control
Hedges' g 0.24 (95% CI 0.02 to 0.46)
2 studies
The review’s own words
weighted mean ES ( g) being 0.24 (95% CI [0.02, 0.46], k = 2)
Affective experience with integrated cognitive and emotional AI support compared with control
Hedges' g 0.42 (95% CI 0.06 to 0.78)
11 studies
The review’s own words
a near-medium effect on affect ( g = 0.42, 95% CI [0.06, 0.78])
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 Scopus, Web of Science, PsycINFO and ERIC on 3 August 2025 and screened 13,879 records. Two researchers reviewed each article independently (Cohen's kappa 0.91 for selection). Studies had to use a controlled or quasi-experimental design comparing AI support with no AI support. Random-effects models pooled Hedges' g.
Of the 54 studies, 17 tested integrated support, 35 tested cognitive-only support and 2 tested emotional-only support. Four extreme outliers from two articles were excluded, including language-skill effects above g = 4. Among randomized trials, 17 of 22 were low risk of bias and 5 had some concerns; among non-randomized studies, 12 of 28 were moderate risk. No study was rated high risk.
Individual results varied widely. A teaching robot gave secondary maths students a large boost in confidence, satisfaction and attention (g = 2.06), and a chatbot in a museum game kept primary pupils engaged longer (g = 1.24). But an interactive agent led secondary students to spend more time gaming the system (g = -1.85).
The authors read results through control-value theory: cognitive support raises students' sense of control, while emotional support steadies achievement emotions. They call for more rigorous experiments to isolate the unique contribution of AI emotional support.
Sources
1 on file- Educational Psychology Reviewdoi:10.1007/s10648-025-10086-4
Talk it over
Who gets this benefit, and who was left out of the study or story?
AI tutors that also support emotions boosted students' knowledge more, a review of 54 studies findshttps://truvace.com/item/emotional-artificial-intelligence-in-education-a-systematic-review-and-meta-anal
Who gets this benefit, and who was left out of the study or story?
The gain A meta-analysis of 54 controlled studies found AI that paired cognitive help with emotional support raised students' knowledge scores by g = 0.88, versus g = 0.52 for AI offering cognitive help alone.
Answer in the threadHow 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 6,151, from Educational Psychology Review.
Answer in the threadWhat does it leave open? What else would you want to know before acting on it?
The record’s own caveat Only two studies tested emotional-only support, and several outcomes pooled very few studies.
Answer in the threadHave 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