When AI gives you feedback, which parts of thinking do you still have to do yourself to actually learn?
When students use AI feedback, which cognitive tasks must they keep doing?
This explores which mental work has to stay with the student, rather than move to the AI, when AI is giving them feedback. The corpus has no study of exactly that setup, so the answer draws on adjacent evidence about learners and AI reliance.
This explores which mental work has to stay with the student, rather than move to the AI, when AI is giving them feedback. The corpus has no study of exactly that setup, so this draws on adjacent evidence about learners and AI reliance. The pattern is consistent: the tasks that build skill are the ones AI makes easiest to hand off, namely attempting, struggling with mistakes, reflecting, and judging.
The first task is making the attempt and wrestling with your own errors. In one study, learners without AI ran into more errors and fixed them on their own, and they kept more skill. Learners who delegated debugging to AI scored lowest on the skill assessment, including the ones who did the most debugging with AI's help (Does AI assistance remove a core learning channel through error work?). Reliance itself can be reduced, though. In a 704-person experiment, feedback that pointed out what offloading was costing people cut requests for answers by half and raised unaided test scores by 51%. A reward for showing effort did nothing measurable (Can metacognitive feedback stop students from offloading to AI?). Telling students what they lose by offloading worked, and paying them to try did not.
The second task is reflecting on your own thinking. In an 80-person study, an AI that asked reflection questions alongside its advice beat an AI that only advised, one that only questioned, and one that did neither (Do reflection questions help people make better decisions with AI?). So the useful feedback keeps the reflection with the student. It asks what they were assuming and leaves the answer for them to work out.
The third task is judging the feedback, and this is the one people underestimate. AI doesn't reduce total task time. It moves time from active work to writing prompts and making sense of outputs (Does AI really save time, or just change how we spend it?). Evaluating becomes the main job, and it is a different skill from producing. It is also easy to do badly. Mistaking a fluent answer for a reasoned one, confusing the model's map with the territory, and seeking confirmation compound when they occur together (Why do people trust AI outputs they shouldn't?). Students need to keep testing feedback against their own understanding instead of absorbing it. They also need to keep the work in their own heads. A four-month EEG study found that heavier LLM use went with weaker brain connectivity, poorer memory retention, and trouble recalling one's own recent work (Does AI assistance weaken our brain's ability to think independently?).
Timing may matter too. Even correct AI suggestions can hurt reasoning by breaking cognitive immersion, so the person has to rebuild focus before continuing (Does AI assistance always help reasoning or does it carry hidden costs?). That note is about reasoning tasks in general, not students, so the classroom version is my inference. The practical rule that follows is to finish a full attempt on your own, then ask for feedback, then do the fixing yourself.