A student aces a task with AI's help — but does that score tell you what they can do alone?

This explores whether a student's score while using an AI tool tells you what they could do alone afterward. The corpus suggests it mostly doesn't, and the two can move in opposite directions. In one study, workers using generative AI did substantially better on content tasks. When they later did similar tasks on their own, they showed no improvement, so the capability never carried over (Does AI assistance help workers learn lasting skills?). Assisted performance measures the human-plus-AI pair, not the person.

The corpus points to a mechanism. Much learning happens when you hit an error and work your way out of it. Learners without AI met more errors and fixed them independently, and they retained more skill. Learners who handed debugging to the AI skipped that work, and the ones who debugged most with AI scored lowest on skill assessments (Does AI assistance remove a core learning channel through error work?). An EEG study over four months found a physical trace of this. Brain connectivity scaled down as AI reliance went up, and heavy LLM users had the weakest neural engagement and the poorest memory of their own recent work (Does AI assistance weaken our brain's ability to think independently?). The more the AI carried the load, the less the student's own head kept.

Students are also badly placed to notice the gap. Polished AI output feels easy to read, and people take that ease as a sign of their own competence, even though they didn't produce the output (Does processing ease mislead users about their own competence?). Researchers call the result the LLM Fallacy: people credit AI output to their own ability, whether or not the output was accurate and whether or not they relied on it heavily. It therefore needs interventions that make the human and machine contributions visible, not just a more accurate model or forced verification (How does AI-assisted work reshape how people see their own abilities?). Self-reports and assisted grades can both mislead.

The gap isn't inevitable, though. In a preregistered experiment with 704 people, feedback that pointed out the cost of offloading cut answer requests to the LLM in half and raised unaided test scores by 51%. A plain effort-based reward did nothing measurable for either outcome (Can metacognitive feedback stop students from offloading to AI?). The better predictor of what a student can do alone is how they used the help, and whether they kept doing the hard part themselves. The only direct check on unaided ability is to test it unaided.