When Students Get Stuck, AI Tutors Should Stop Asking Questions

AI tutors are now common in schools and study apps, and most are designed with one rule: never just hand over the answer. Instead, they nudge students with questions like "what do you think happens next?" That approach is meant to encourage real thinking. But researchers wanted to know what happens when a student is truly stuck — not thinking productively, just spinning their wheels.

To find out, they analyzed 20,462 student replies from 1,260 real tutoring sessions with an AI chemistry tutor. They identified 6,630 "impasse" moments — times when a student made a conceptual error, said they were confused, or asked for help. The pattern was stark: each additional stuck back-and-forth lowered the odds the student would recover on the next turn by 12.7%. In other words, the deeper the hole, the harder it is to climb out. Some students quit before they ever got to try a problem.

The team also tested three tutoring styles. A baseline tutor gave the answer directly about half the time. A strict "never give the answer" tutor asked a follow-up question every single time. A guided tutor adapted. The adaptive one won. Asking a good question helps early on, but its value fades as the student stays stuck. Repeating a question that already failed led to recovery only 28% of the time, compared with nearly 40% when the tutor simply addressed the student's mistake.

The takeaway is that good tutoring is about timing, not rules. Software can already spot signs of being stuck — confusion, repeated wrong answers, help requests — and could switch from asking to explaining. The catch: this was one subject, one tutor design, and partly simulated conditions, so it's a strong signal rather than a final answer. The point isn't to hand out answers immediately; it's to notice when a student has struggled enough and finally step in.