AI Access, Student Learning, and Academic Achievement Inequality

Last registered on September 25, 2026
View Trial History

Pre-Trial

Trial Information

General Information

Title
AI Access, Student Learning, and Academic Achievement Inequality
RCT ID
AEARCTR-0019769
Initial registration date
September 19, 2026

Initial registration date is when the trial was registered.

It corresponds to when the registration was submitted to the Registry to be reviewed for publication.

First published
September 25, 2026, 10:06 AM EDT

First published corresponds to when the trial was first made public on the Registry after being reviewed.

Locations

There is information in this trial unavailable to the public. Use the button below to request access.

Request Information

Primary Investigator

Name
Chaonuo Dai
Affiliation
City University of Macau
Contact Primary Investigator

Other Primary Investigator(s)

Additional Trial Information

Status
On going
Start date
2026-08-24
End date
2026-12-31
Keywords
Education, Labor
Additional Keywords
generative artificial intelligence, student learning, academic achievement, human capital, higher education, achievement inequality
JEL code(s)
I21, I23, I24, J24, C93
Secondary IDs
Prior work
This trial does not extend or rely on any prior RCTs.
Abstract
This randomized controlled trial examines whether generative artificial intelligence (AI) access during learning improves academic performance and changes achievement gaps by prior preparation. Conducted in an Introductory Econometrics course at City University of Macau with approximately 120 third-year undergraduates, it includes consenting enrolled students. After a supervised diagnostic test and collection of pre-treatment GPA, participants are individually randomized in roughly equal proportions through blocked randomization to fixed AI-access or control groups. Both groups complete identical supervised weekly tasks with equal time allowances and non-AI materials; only the treatment group may use AI. Common closed-book midterm and final examinations prohibit AI. Primary analyses estimate the average assignment effect on standardized final-examination scores and its interaction with baseline diagnostic achievement. Secondary analyses examine midterm scores, weekly performance, and examination-score dispersion. The study distinguishes AI-assisted performance from unaided learning and assesses whether initially lower- or higher-achieving students benefit more.
External Link(s)

Registration Citation

Citation
Dai, Chaonuo. 2026. "AI Access, Student Learning, and Academic Achievement Inequality." AEA RCT Registry. September 25. https://doi.org/10.1257/rct.19769-1.0
Sponsors & Partners
Experimental Details

Interventions

Intervention(s)
The treatment is access to generative artificial intelligence (AI) during 12 weeks of in-class econometrics practice.
Intervention Start Date
2026-08-31
Intervention End Date
2026-12-31

Primary Outcomes

Primary Outcomes (end points)
1. Scores from 12 weeks of in-class econometrics practice, completed under the assigned AI-access or control condition.
2. Score on the common closed-book midterm examination, taken without AI.
3. Score on the common closed-book final examination, taken without AI.
Primary Outcomes (explanation)
For each of the 12 practice weeks, student answers are marked using that week’s reference answers and scoring rubric, with the same criteria applied to both groups. The weekly score measures performance while students work under their assigned condition. Practice difficulty and maximum marks may differ across weeks.
The midterm and final examinations use common questions and marking criteria for both groups. Because AI is prohibited in both examinations, their scores measure learning demonstrated without AI after intermediate and full-course exposure, respectively. Higher scores indicate better performance on every assessment.
Raw scores and maximum possible marks are retained. For analysis, each practice score is standardized separately for that week using the pooled mean and standard deviation across both arms among students with observed scores. Midterm and final scores are standardized separately in the same way. The 12 weekly scores enter a repeated-observation analysis; midterm and final scores are analyzed separately. All three performance domains are primary.

Secondary Outcomes

Secondary Outcomes (end points)
AI-use behavior among treated students, measured from their complete submitted dialogues. Measures describe the intensity of interaction, the purposes for which AI is used, and the extent of follow-up engagement.
Secondary Outcomes (explanation)
The unit of coding is the student’s dialogue record for a practice session. Interaction intensity is measured by the number of student prompts. Requests are coded by purpose, including generating answers, obtaining explanations, checking an answer, and correcting an error. Categories may overlap. Follow-up engagement records whether students ask additional questions, request clarification or challenge an AI response after the initial exchange. An indicator identifies sessions containing only one student prompt.
These measures are summarized by student and week to describe AI-use patterns and their evolution. Coding rules will be documented, and only behavior supported by the submitted dialogue will be recorded. Unsubmitted or incomplete dialogues are flagged; absence of a record is not treated as evidence of no AI use. Dialogue-derived behavior is measured within the treatment group. Associations between particular usage patterns and scores are descriptive because usage choices are not randomized.

Experimental Design

Experimental Design
This individually randomized trial recruits approximately 120 consenting third-year students from an undergraduate Econometrics course at City University of Macau. Students are assigned in approximately equal proportions to unrestricted generative artificial intelligence (AI) access or no AI during 12 weeks of lecturer-monitored in-class practice. Primary outcomes are practice, midterm and final-examination scores; both examinations prohibit AI. The secondary outcome is AI-use behavior measured from submitted dialogues.
Experimental Design Details
Not available
Randomization Method
Randomization is conducted by computer separately within each participating class, after baseline measurement and before treatment tasks begin. Within each class, students are randomly assigned in approximately equal numbers to AI access or control. For classes with an odd number of students, the additional allocation is determined randomly. Class is the randomization block. Baseline achievement and pre-treatment grade point average are included as analysis covariates but are not additional randomization strata. The allocation code and resulting assignments are retained. Assignment remains fixed throughout the study.
Randomization Unit
Individual student.
Was the treatment clustered?
No

Experiment Characteristics

Sample size: planned number of clusters
Not applicable.
Sample size: planned number of observations
Approximately 120 students in total. Repeated weekly task records are observations on these same students, not additional participants. In total, we may have aroun
Sample size (or number of clusters) by treatment arms
Approximately 120 students in total. Repeated 12 weekly task records and 2 exam test scores are observations on these same students. In total we may have around 1600 observations. Attrition may happen if student skip the class or miss the midterm or final exams.
Minimum detectable effect size for main outcomes (accounting for sample design and clustering)
The planned sample is approximately 120 students. As an illustrative calculation, 60 students per arm provide approximately 80 percent power to detect an average effect of 0.51 outcome standard deviations on a single examination with a two-sided 5 percent test, using a normal approximation, equal arm variances, no attrition and no precision gains from baseline adjustment. This calculation does not incorporate adjustment for multiple primary outcomes. Power for the weekly practice panel additionally depends on within-student score correlations. Power for baseline-achievement interactions is not established by this calculation. Attrition reduces power; baseline adjustment may improve precision.
Supporting Documents and Materials
IRB

Institutional Review Boards (IRBs)

IRB Name
Human Research Ethics Committee for Non-Clinical Faculties
IRB Approval Date
2026-09-17
IRB Approval Number
2626-RE-64
Analysis Plan