Teacher-Mediated Generative AI use and Student Learning: A Field Experiment on Adaptive Revision in Government Secondary Schools in Rajasthan, India

How to cite

Bhardwaj, Sajal; Bhutoria, Aditi. 2026. "Teacher-Mediated Generative AI use and Student Learning: A Field Experiment on Adaptive Revision in Government Secondary Schools in Rajasthan, India." AEA RCT Registry. AEARCTR-0018900. https://doi.org/10.1257/rct.18900-2.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Sajal Bhardwaj* , Aditi Bhutoria

* Primary investigator

Registry ID
AEARCTR-0018900
DOI
https://doi.org/10.1257/rct.18900-2.0
First registered
Last updated
Status
Completed
Trial dates
2025-12-01 – 2026-03-10
Countries
India (Jaipur, Rajasthan)
JEL codes
C93, I21, I24, O33
Registry record
http://www.socialscienceregistry.org/trials/18900

Abstract

This study reports a school-randomised field experiment on whether teacher-mediated, generative-AI-supported revision improves student learning in resource-constrained government schools. From an eligible pool of schools provided by the Ministry of Education, Government of Rajasthan, we randomised 43 schools to treatment (22) and control (21). In treatment schools, teachers used a curriculum-aligned custom GPT system on classroom smartboards during pre-examination revision in mathematics and science for Classes 7 and 9. At the beginning of each session, the system offered three modules (concept revision, practice questions, and doubt resolution) and generated responses live at three difficulty levels (basic, intermediate, and advanced), enabling within-classroom differentiation that conventional revision strategies cannot easily provide at scale. The usage was teacher-mediated, at an intended dosage of two to three 30-minute sessions per week. Control schools conducted equivalent teacher-led revision without technology.

We conceptualised the intervention as AI-enabled adaptive instructional scaffolding and differentiated instruction. The system configured explanations and practice to students' learning levels and could increase behavioural and emotional engagement, reduce disaffection, and thereby improve academic achievement. The primary outcome is academic achievement, measured by a curriculum-aligned test. Secondary outcomes are academic procrastination (González-Brignardello & Paniagua, 2023) and four engagement/disaffection subscales (Skinner et al., 2008), measured at baseline and endline. We also conduct an exploratory heterogeneity analysis to test whether treatment impacts baseline present bias. The logic being, students who experience studying as more effortful may benefit disproportionately when adaptive support lowers the immediate cost of revision.

After randomisation but before baseline, seven treatment schools withdrew because teachers were assigned to Booth Level Officer duties for cyclical electoral roll revision. This exogenous shock was unrelated to treatment assignment or school characteristics. The analysis sample is 15 treatment and 21 control schools (1,126 students at endline). We assess the administrative balance between retained and withdrawn schools and document the incidence of electoral duty across both arms. Estimation uses ANCOVA, with robustness checks via difference-in-differences, Lee bounds, and the Romanov-Wolf test. This study was registered after data collection; its hypotheses are exploratory rather than prespecified.

Intervention

The treatment schoolteachers used a custom GPT aligned with the curriculum requirements of classes 7th and 9th. They used the tool on their smartboards for revising the mathematics and science syllabus before the annual examinations. At the beginning of every session, the tool presented three modules: concept revision, practice questions, and doubt resolution. Upon selecting a module, it asked teachers to specify the unit/topic they wanted to revise or the practice questions they wanted to provide. Upon specification, the tool generated live responses at three levels of difficulty- basic, intermediate, and advanced; this enabled a within-classroom differentiation of explanations and practice. The students did not interact with the system directly; teachers mediated the sessions. The intended dosage was 3 times a week, with 30-minute sessions each. Control schools conducted equivalent teacher-led revision without any technological support.

Experimental design

A two-arm, parallel, cluster-randomised field experiment. Schools were drawn from an eligible pool provided by the Ministry of Education, Government of Rajasthan, and were randomly assigned to the treatment or control group. Treatment schools delivered teacher-mediated, GPT-supported adaptive revision while control schools delivered teacher-led revision without technology. We collected outcomes at the student level at baseline and endline.

Design details

Randomization method
Randomised using a computer
Randomization unit
Cluster randomised at the school level.
Clusters
43 schools (assigned: 22 treatment, 21 control). Realised analysis sample following exogenous attrition: 36 schools.
Observations
Baseline: 1395 students Endline: 1126 students
Arms
Randomised 22 schools into treatment and 21 into control groups (total 43 schools). Analysis sample: 15 treatment, 21 control (36 schools), following exogenous attrition of 7 treatment schools to Booth Level Officer electoral duties.
Primary outcomes
Academic achievement in mathematics and science, measured by a curriculum-aligned assessment administered at baseline and endline.
Secondary outcomes
Academic Procrastination, Behavioural engagement, Emotional engagement, Behavioural disaffection, and Emotional disaffection. All these outcomes were measured both at baseline and endline.

Cited by

No citation data available for this registration yet.

Keywords