Can AI Tutors Teach Struggling Readers? Experimental Evidence from Secondary Schools in Colombia

How to cite

Rozo, Sandra. 2026. "Can AI Tutors Teach Struggling Readers? Experimental Evidence from Secondary Schools in Colombia." AEA RCT Registry. AEARCTR-0019168. https://doi.org/10.1257/rct.19168-1.0

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

Sandra Rozo*

* Primary investigator

Registry ID
AEARCTR-0019168
DOI
https://doi.org/10.1257/rct.19168-1.0
First registered
Last updated
Status
In development
Trial dates
2026-08-01 – 2027-01-31
Countries
Colombia (Barranquilla)
JEL codes
I21, O33, I25, I28
Registry record
http://www.socialscienceregistry.org/trials/19168

Abstract

Many secondary students in developing countries cannot read fluently enough to learn from written text. This study evaluates whether an artificial intelligence reading tutor improves oral reading fluency and reading comprehension among struggling readers in public secondary schools in Barranquilla, Colombia. We randomly assigned 31 classes serving 845 students in 11 schools to a structured remedial reading program (Avanza+), or to the same program with its oral reading routines completed on Microsoft Reading Coach, which scores speed and accuracy through speech recognition and generates individualized practice on the words each student misses. Because both groups receive identical curriculum and instructional time, the design identifies the marginal contribution of automated, individualized feedback the alternative for education systems considering AI tutors as a complement to structured remediation.

Intervention

All students in the study receive Avanza+, a structured remediation program developed by the World Bank for secondary students who can decode but have not automatized reading. The program includes 48 one-hour sessions, delivered three times per week over approximately 16 weeks. Sessions are grouped into eight units and combine explicit instruction in phonological awareness, spelling, morphology, and vocabulary with repeated oral reading routines and formative assessments. Treatment classes implement the identical curriculum with one difference: the oral reading routines, consisting of three timed readings per unit and the end-of-unit fluency check, are completed on Microsoft Reading Coach using tablets/computers. In each exercise, the student reads aloud for one minute; the platform scores speed and accuracy through speech recognition, returns immediate feedback, and generates individualized practice on the words the student misread. Control classes complete the same routines on paper, working in pairs: one student reads aloud while the other times the reading and provides feedback, with occasional support from the teacher. The treatment effect therefore captures the marginal value of automated, individualized feedback relative to peer feedback, holding curriculum and instructional time constant.

Experimental design

The experimental sample includes 845 students in 31 instructional groups across 11 public schools in Barranquilla. The unit of randomization is the instructional group. Randomization was blocked by school: within each school, half of the groups were randomly assigned to treatment, and in schools with an odd number of groups a fair coin determined the extra assignment.

Design details

Randomization method
computer randomization
Randomization unit
Instructional group (classroom).
Clusters
The study was blocked by schools and clustered by instructional class.
Observations
842 students
Arms
842 students, 11 schools, and 31 clusters
Primary outcomes
The study has two primary outcomes. The first is oral reading fluency, measured as words read correctly per minute (WCPM) on a one-minute reading of a 308-word informative text. The second is reading comprehension, measured as the number of correct answers (0 to 10) on a multiple-choice test with six literal and four inferential items based on a second informative text.

Cited by

No citation data available for this registration yet.

Keywords