When students learn with an AI tutor, what separates using it to think from using it to skip the thinking?

This explores which ways of working with an AI tutor or chatbot let students actually learn the material, rather than just finish the task with AI help. The corpus has one direct classroom study and a set of indirect findings, and it doesn't yet contain a proven recipe. What it does have is a pattern: learning survives when the AI keeps the student doing the thinking, and it erodes when the AI hands over a finished product.

The closest direct evidence is a study where students worked with a chatbot instead of with peers. They did better on practical tasks and produced more knowledge-based dialogue, meaning they explained and elaborated on concepts. The cost was that they spoke far less overall and expressed far fewer personal perspectives (Does chatbot interaction trade authenticity for better problem-solving?). So the chatbot pattern preserved, and even improved, the knowledge side of learning while thinning out the part where students voice their own views. Whether that trade is acceptable depends on what the course is for.

The warning signs come from the other direction. AI can now produce the outward form of intellectual work, such as an essay, a proof or a code solution, without the reasoning that normally produces it (Does AI separate intellectual form from the thinking behind it?). In a classroom this means a polished answer no longer shows that anyone learned anything. It gets worse because fluent AI output invites trust it hasn't earned. Confusing a plausible-sounding answer with a correct one, and reading confirmation of your existing belief as evidence, tend to reinforce each other (Why do people trust AI outputs they shouldn't?). The patterns most likely to protect learning are therefore the ones that make students check, explain and reason, not the ones that let them copy.

The research also suggests looking at how students interact, not only at what they turn in. In coding-agent conversations, the way people talked to the AI explained their outcomes beyond what their prior skill predicted. But those traits weren't stable or transferable enough to count as a teachable skill (Can conversation patterns predict coding outcomes better than prior skill?). That means we can spot good and bad interaction patterns after the fact but can't yet tell students to do X and expect it to work. This fits a wider move in agent research, where evidence is shifting from final answers to whole interaction trajectories (How should we evaluate agent behavior beyond final answers?). Assessing learners by their process, not just their output, follows the same logic.

Two more findings point at what a learning-friendly AI might do, though neither was tested in a classroom. Signals like hesitation, gaze and typing speed can reveal when someone is struggling, so a tutor could step in at the right moment without interrupting with quizzes. The same signals also enable manipulative profiling (Can AI systems read cognitive state from interaction patterns alone?). And in model training, teacher-refined data hurts a student model when it sits beyond what that student can currently absorb, even if the data is objectively better (Does teacher-refined data always improve student model performance?). That result is about machines, but it hints that help pitched above the learner's level may be no help at all.