Overview
In real K–12 settings, AI tutoring is promising when access becomes repeated, productive practice, timely reasoning support, and independent performance later. Evidence from literacy support trials, the two-year Khanmigo program, and the NUMI experiment locates different bottlenecks: human facilitation can raise participation, structured routines can raise practice, and embedded AI can improve recovery after mistakes at a time cost. They support conditional deployment and delayed measurement rather than a universal claim about AI alone or long-run effects.
Learning outcomes
- Access is necessary but does not guarantee recurring use, productive practice, dialogue, or durable learning.
- Human facilitation increased literacy-platform engagement in two trials, but the added dose did not improve end-of-year reading achievement.
- The two-year Khanmigo program increased structured mathematics practice and achievement, but its bundled design does not isolate the AI tutor’s contribution.
- NUMI found more accurate error recovery with embedded AI support, but students spent more time and showed only a modest, localized delayed-learning signal.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
Why is access alone insufficient for AI tutoring?
Access alone is insufficient because students may not start, return, practice productively, seek help, or retain what they do inside the platform. These transitions can fail separately, so availability is not a learning outcome.
Why should schools separate practice from tutoring dialogue?
Practice and tutoring dialogue are separate because a student can complete structured work without asking the AI for substantive help. Khanmigo increased practice in a supervised routine, while the median matched Year-2 student messaged during only 17% of mistake sessions.
What can human support add to an AI-tutoring program?
Human support can add accountability, motivation, troubleshooting, group norms, and reflection, increasing platform engagement without providing the subject instruction. In the literacy trials, the added use remained too small or inconsistent to improve end-of-year reading achievement.
What evidence shows that AI tutoring improved learning?
Evidence of improved learning requires later independent performance, not just access, messages, completion, or a platform-defined streak. Delayed tests should include practiced and, where possible, unpracticed or transfer material, and short follow-ups remain limited.