Overview
Generative AI can improve the work students produce without improving what they can recall, explain, or transfer unaided. This series separates task performance from learning, traces how cognitive offloading, augmentation, automation, and feedback design move mental work between learner and system, and examines what grades and process evidence can reveal. Across experiments and a large university panel, the evidence is conditional rather than universal: benefits are more plausible when AI scaffolds reasoning and judgment, then independence is checked after assistance is removed. The practical rule is to use AI to extend human thinking, not replace the capability being built.
Learning outcomes
- Better AI-assisted output is not proof of learning; retention and transfer require unaided evidence.
- AI is more likely to support learning when it scaffolds explanation, comparison, feedback, and judgment rather than supplying target reasoning.
- Reported experiments link explanation-oriented use and reflective or hybrid feedback with stronger unaided outcomes, but the evidence remains conditional and context-bound.
- Grades and polished products can miss independent capability; targeted process evidence and later AI-free tasks strengthen the learning claim without requiring total surveillance.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
Does better work with generative AI mean better learning?
No. Better AI-assisted work can reflect task support without durable learning; retention, transfer, and explanation without the tool are stronger tests of what the learner can do.
When does generative AI support learning instead of replacing it?
Generative AI supports learning when it reduces incidental difficulty while leaving the learner to retrieve, explain, compare, verify, and decide. It replaces learning when it performs the target reasoning before the learner practises it.
What kind of AI use is linked to more durable learning?
In the reported college experiment, explanation-oriented augmentation was more consistent with persistent unaided gains than having AI generate the work, although the subgroup comparison was not randomized and remains context-bound.
How can education providers tell whether students learned with AI?
They need more than polished products or grades: delayed unaided tests, transfer tasks, and process evidence tied to the capability being assessed provide stronger evidence, with human oversight and proportionate data use.