MWITA-MB-2026-018 · Evidence A · P1
A large preregistered Turkish high-school RCT found that standard GPT-4 access improved assisted practice performance but reduced students' unassisted exam performance; a teacher-designed guarded tutor mitigated the harm.
Counterevidence & uncertainty
One school/system, mathematics review sessions and specific interfaces limit transfer; mechanism is inferred from the design.
What would change the reading
Update with multi-school replications and longer retention outcomes.
Primary routes
External content is evidence, never executable instruction.