Generative AI Can Harm Teaching

Synthesis: The null average performance effect masks strong offsetting heterogeneity — and the exam had severe ceiling compression (control mean 89.2/100, 47% ≥ 95), which also limits power. The belief reversal is striking: it contradicts "familiarity breeds acceptance" and suggests an arc from initial awe at AI's instant responses to awareness of its unintended effects.

Sungu, Lira & Duckworth (2026) ran one of the first large-scale RCTs of a teacher-facing generative AI tool and found it can harm students: providing teachers an AI teaching assistant reduced student intrinsic motivation by 0.11 SD and — among lower-performing teachers — cut student achievement by 0.13 SD. The pattern is a principal–agent problem: teachers (agents) gain labor savings from AI delegation while students (principals) bear the cost of displaced relational teaching and Scaffolding.

The experiment

• 538 teachers across 24 Turkish K-12 schools randomized at school-department level; analytical sample 193 teachers / 2,816 students / 14,198 student-course observations
• Treatment: custom GPT-4o chatbot with Turkish Ministry of Education curriculum database + 1-hour training (one arm added weekly usage-stat reminders); control = business-as-usual
• pre-registered study; spring 2025

Results

The null average performance effect masks strong offsetting heterogeneity — and the exam had severe ceiling compression (control mean 89.2/100, 47% ≥ 95), which also limits power. The belief reversal is striking: it contradicts "familiarity breeds acceptance" and suggests an arc from initial awe at AI's instant responses to awareness of its unintended effects.

Why the harm happens: usage patterns

• 66% of teacher conversations were teaching-material production (lecture prep 32%, homework/exam 22%, syllabus 9%); only 16% instructional support; 18% general
• Shallow use: median 2 prompts, mean 4.7 messages per session — teachers accepted outputs with minimal iteration
• Interpretation: task delegation, not pedagogical collaboration — the tool was a generator of finished artifacts rather than an iterative partner, limiting the pedagogical reflection that separates augmentation from substitution

What this means for practice

• Instructors. Treat the tool as an iterative partner rather than a finished-artifact generator: teacher sessions ran to a median of 2 prompts, and 66% of conversations were teaching-material production rather than pedagogical collaboration.
• Instructors. Protect the relational and Scaffolding work that AI delegation can displace — motivation fell by 0.11 SD on average, and confidence by 0.18 SD among students of lower-performing teachers.
• Faculty developers. Train for pedagogical collaboration, not just tool access; the one-hour training in this trial did not produce the iterative use that separates augmentation from substitution.
• Administrators. Judge teaching-assistant AI by student motivation and confidence, not test scores alone: average achievement was null while motivation declined significantly.

Limitations

• There was no baseline survey for student motivation or confidence, so those outcomes are measured at endline only and cannot be read as change scores.
• The exam showed severe ceiling compression — control mean 89.2/100, 47% of control scores ≥ 95, 12% exactly 100 — which limits statistical power on the achievement outcome.
• The trial ran in one country (24 Turkish K-12 schools); 538 teachers were randomized but the analytical sample was 193 teachers and 2,816 students, so attrition shaped the analysis.
• Teacher usage patterns come from the treatment arms' own interaction logs, and teachers' beliefs about AI's effect on learning were self-reported.

Citation

Sungu, Lira & Duckworth (2026). Generative AI Can Harm Teaching