Enhancing School Students’ Self-Regulated Learning through Generative AI Support: A Randomized Controlled Trial
Lower secondary (grades 7–9) AI and learning Foundations of optimal learning
Links to claims
Supports (1)
| Claim | Basis |
|---|---|
| C5 | Grades 7–9 preregistered RCT (Germany): GPT-4o with Socratic strategy prompts or utility-value prompts did not raise tested domain knowledge over a plain answer-withholding chatbot across four lessons (χ²(1) = 1.28, p = .257); consistent with C5's bound that LLM tutors have not been shown to reach step-based tutoring levels. Short, immediate test, no non-AI control. |
Review state: first-pass draft, pending human review.
Source note
Syntes: Fütterer et al. (2026) — pedagogically prompted GPT-4o did not raise pupils' knowledge over a plain chatbot (grades 7–9)
Källa: Fütterer, T.; Bardach, L.; Kuhn, J.; Keller, S. D.; Gerjets, P. (2026 (Educational Psychology Review 38, article 42; published 2026-04-14)). Enhancing School Students’ Self-Regulated Learning through Generative AI Support: A Randomized Controlled Trial. https://doi.org/10.1007/s10648-026-10133-8
Verified numbers (full text read 2026-10-11, Springer open access)
- Sample: N = 371 (mean age 13.9; 54 % female), grades 7–9, German secondary schools; 130 (35 %) dropped out between pretest and posttest, not related to baseline characteristics.
- Conditions: (a) GPT-4o utility-value reflection, (b) GPT-4o cognitive-strategy prompting (explain, reflect, self-correct; Socratic questioning), (c) control: GPT-4o that chatted about the task without pedagogical strategies. All three were instructed not to give correct answers; solutions were shown on the next page afterwards.
- Domain knowledge (H5): condition × time not significant, χ²(1) = 1.28, p = .257; no main effects; robust across imputation, completers, covariates and complier analysis.
- Strategy use: self-reported elaboration χ²(2) = 0.08, p = .960; applied learning-by-explaining χ²(2) = 0.01, p = .992. Utility value: the utility-value condition did better than the strategy condition (β = −0.50), not better than control.
- Exploratory: pupils who interacted more meaningfully had higher posttest knowledge (β = 0.19, p = .003) regardless of condition; teachers reported some pupils mainly sought correct answers.
- Limits: short (four learning sessions), only three knowledge items per subject, no non-AI control group, no delayed test.
Why it is in the library
A preregistered K–12 classroom RCT showing that prompt-engineered "tutor" behaviour (Socratic, answer-withholding) did not by itself change tested knowledge compared with a plain chatbot over four lessons. It bears on C5's bound about LLM tutors and on Q2 (what AI tools in Swedish-type classrooms actually do).
Metadata as recorded
- ID
- AI-025
- Tier
- RCT
- Closed-book follow-up
- no delay (posttest directly after the four learning sessions; tested without the chatbot)
- Titel
- Enhancing School Students’ Self-Regulated Learning through Generative AI Support: A Randomized Controlled Trial
- Författare
- Fütterer, T.; Bardach, L.; Kuhn, J.; Keller, S. D.; Gerjets, P.
- År
- 2026 (Educational Psychology Review 38, article 42; published 2026-04-14)
- Källa/DOI
- https://doi.org/10.1007/s10648-026-10133-8
- URL
- https://link.springer.com/article/10.1007/s10648-026-10133-8 (open access); preregistration https://doi.org/10.17605/OSF.IO/VBSQ4
- Typ
- preregistered individually randomised trial, 371 pupils in grades 7–9 (Gymnasium and Realschule, Baden-Württemberg), six 45-minute sessions in regular physics or English lessons (pretest, four learning sessions, posttest), spring 2025; three GPT-4o conditions: utility-value reflection, cognitive-strategy (Socratic explaining) prompting, control = plain chatbot without pedagogical prompting; all chatbots told not to give answers
- Åldersgrupp
- grundskola-hog
- Teman
- ai-och-larande;optimalt-larande-grunder
- Hämtad
- 2026-10-11
- Paywall
- open access
- Tillagd
- source sweep 2026-10-11 (kind: new); links pending human review
- Full text (fetched 2026-10-16, legal open-access copy, private)
- fulltext.pdf + — via unpaywall: https://link.springer.com/content/pdf/10.1007/s10648-026-10133-8.pdf
The original files (PDF or abstract) are kept in the project archive and are not published here.