evidoria

← Back to browse

Good practice Imported

Generative AI Can Harm Teaching — A 14-School RCT of a ChatGPT Teaching Assistant in Turkey

Türkiye · Istanbul · See the Türkiye profile · See the Istanbul profile

Evidence: Randomised controlled trial Top 77% 40/100 · Ask Evidence Copilot about this practice

A randomized trial of 193 teachers and 2,816 students across 14 Turkish schools found a GPT-4o teaching assistant left average achievement unchanged but significantly lowered both achievement and confidence among students of already lower-performing teachers, and made classes fee

14 schools
Schools in the trial
2816 students
Students in the trial
-0.129 SD
Achievement change for students of lower-performing teachers
-0.11 SD
Overall change in student intrinsic motivation

Details

Promoter
University of Pennsylvania (Wharton School / Duckworth Lab)
Period
Spring 2025 (10-week RCT); preprint released June 2026
Keywords
K-12 education, generative AI, teacher tools, randomized controlled trial

Context

Researchers from the University of Pennsylvania's Wharton School (Alp Sungu, Benjamin Lira and Angela Duckworth) ran a randomised controlled trial across a private K-12 school network in Turkey during spring 2025, testing a custom GPT-4o-based teaching assistant built on the Turkish Ministry of Education curriculum.

Objectives

The trial aimed to measure the causal effect of giving teachers access to a curriculum-aligned generative-AI teaching assistant on student academic achievement, confidence and motivation.

Activities

193 teachers and 2,816 students (14,198 student-course observations across 14 schools) were randomised to business-as-usual teaching, AI-tool access, or AI-tool access plus weekly usage reminders. Teachers used the assistant mainly for lesson materials, assessments and administrative communication over a 10-week period.

Results

Average academic achievement was unchanged overall, but among students of teachers who were already lower-performing, both achievement (-0.129 SD) and confidence fell significantly, and treatment-group students rated classes as less enjoyable, less interesting and less important. Student intrinsic motivation fell by 0.11 SD overall, roughly three times more among teachers who were already heavy AI users pre-study.

Conclusions

The study finds that generative-AI access can harm teaching outcomes for students of already lower-performing teachers and reduce student motivation overall, even though average achievement was unaffected; the study is a working paper (SSRN, posted 25 June 2026) that has not yet completed peer review.

Implementation
Indicative cost
Medium (€50k–€500k)
Time to results
Short (< 1 year)
Staffing & skills
University of Pennsylvania Wharton School / Duckworth Lab (Alp Sungu, Benjamin Lira, Angela Duckworth)

Conditions for success

  • AI tool built on the Turkish Ministry of Education curriculum
  • Random assignment across three arms: business-as-usual, AI access, and AI access plus weekly usage reminders

Common failure modes

  • Unrestricted AI-tool access lowered achievement (-0.129 SD) and confidence among students of already lower-performing teachers
  • Overall student intrinsic motivation fell 0.11 SD, about three times more among teachers who were already heavy AI users pre-study
  • Treatment-group students rated classes as less enjoyable, less interesting and less important
  • Findings come from a working paper that has not yet completed peer review

Replication kit

Reusable artefacts from this practice — as published by their sources.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Similar practices you may find useful