An NBER trial of 6,997 US middle schoolers found a guard-railed AI math tutor improved post-mistake recovery but slowed practice, with modest delayed gains when paired with mastery.