A general model becomes a tutor only inside structured pedagogy.
Executive brief
The model made supervised practice abundant; it did not replace the teacher.
AI value · Endline assessment
Verified0.310 SD weighted total score; 0.238 SD English skills (abstract reports 0.23; narrative rounds to 0.24)
World Bank Policy Research Working Paper, not formal peer review; 654 observations in the main outcome tables; short, supervised intervention among interested students.
Before
Continues business-as-usual classes and study without the six-week after-school intervention. → Delivers ordinary curriculum and grades school assessments.
After
Completes curriculum-aligned English, AI, and digital-skills activities in twelve 90-minute lab sessions. → Introduces activities, monitors model use, and redirects shortcutting or unsuitable responses.
Human boundary
Teachers own pedagogy and monitoring.
Why it matters
A general model becomes a tutor only inside structured pedagogy.
Before
How the work ran before the change.
Public-school student
Continues business-as-usual classes and study without the six-week after-school intervention.
ControlRandomized control condition; the paper does not claim a separate prior tutoring queue.
Classroom teacher
Delivers ordinary curriculum and grades school assessments.
ControlHuman pedagogy and assessment authority.
What changed
A general model becomes a tutor only inside structured pedagogy.
Decision rightHuman moves from creator to judge
After
How the same work runs now.
Student pair using Microsoft Copilot/GPT-4
Completes curriculum-aligned English, AI, and digital-skills activities in twelve 90-minute lab sessions.
ControlCarefully crafted prompts; pair work; intermittent connectivity/electricity are documented failure modes.
Teacher facilitator
Introduces activities, monitors model use, and redirects shortcutting or unsuitable responses.
ControlTeacher retains pedagogy; outcomes measured independently after six weeks.
Exception path
Internet, electricity or unsuitable output is handled by teacher.
Decision authority
Teachers own pedagogy and monitoring.
| # | Actor | Action | Control |
|---|---|---|---|
| 01 | Public-school student | Continues business-as-usual classes and study without the six-week after-school intervention. | Randomized control condition; the paper does not claim a separate prior tutoring queue. |
| 02 | Classroom teacher | Delivers ordinary curriculum and grades school assessments. | Human pedagogy and assessment authority. |
| # | Actor | Action | Control |
|---|---|---|---|
| 01 | Student pair using Microsoft Copilot/GPT-4 | Completes curriculum-aligned English, AI, and digital-skills activities in twelve 90-minute lab sessions. | Carefully crafted prompts; pair work; intermittent connectivity/electricity are documented failure modes. |
| 02 | Teacher facilitator | Introduces activities, monitors model use, and redirects shortcutting or unsuitable responses. | Teacher retains pedagogy; outcomes measured independently after six weeks. |
Work that left the path
Human role before
Classroom teachers delivered the ordinary curriculum and governed school assessment; the randomized control continued business-as-usual without a separate AI tutoring queue.
Human role after
Teachers facilitate, monitor and prevent shortcutting; assessments remain governed.
AI role
Provides interactive practice through curriculum-aligned prompts.
Control business-as-usual0.310 SD weighted total score; 0.238 SD English skills (abstract reports 0.23; narrative rounds to 0.24)
Six weeks; 12 sessions · Nine public schools
World Bank Policy Research Working Paper, not formal peer review; 654 observations in the main outcome tables; short, supervised intervention among interested students.
Anti-pattern
Do not generalize to unsupervised access.
Questions
Portability conditions
Reputation risk
low