AI that asks instead of answers
AI can make practice look better while making learning worse. So the AI in The Learning Library is built to ask questions, not hand over answers.
The warning
In a study of high school maths students in Turkey, practice grades rose 48% with plain GPT-4 and 127% with a version built to tutor. But once access was taken away, the plain-GPT-4 group scored 17% below students who never had it. The tutoring guardrails largely prevented that drop (Bastani et al., 2025, working paper).
Guided AI tutoring can work well. In a Harvard physics trial, students' median learning gains with a structured AI tutor were more than double those from in-class active learning (Kestin et al., 2025). Feedback on each step comes close to a human tutor, while feedback on just the final answer helps far less (VanLehn, 2011).
The difference is whether the AI does the thinking for the student.
What we built
- Explain it back. Students explain an idea in their own words. The feedback names what they have right, says which idea is missing without stating it, and asks one guiding question.
- Teach it. Students teach Sam, an AI classmate who missed the lesson. Sam asks what a curious beginner would, aimed at whatever hasn't been explained yet, and never fills in the answer.
- An answer-leak guard. If feedback repeats eight or more words of the source explanation, that part is replaced before the student sees it.
- Giving up is allowed, and it's honest. Students can see the model answer, but it counts as not recalled.
- Check-ins without AI. The questions we use to judge what's working have no hints and no AI help, so we always know what students can do on their own.
How this connects to our solution
- Our solution, "AI that helps you think": this post is the evidence behind that card.
- Belief 8: "Teachers stay in charge, and less is more. AI drafts; teachers choose, swap and edit."
- In the product: Explain it back, Teach it, and "Why is it true?" in Learn.
- Call to action: See what we believe
