This post from The Learning Scientists looks at recent research on AI and learning and asks why tools that make work easier can leave students worse off later. The author reads 2 new studies alongside older findings from memory research.
In a study of Chinese students in grades 7 to 12, heavy generative AI users scored well on homework but worse on exams, while students who skipped AI or used it strategically showed the reverse. In a second study, half the participants had an AI assistant for 12 fraction problems, then lost it without warning for 3 more. The other half solved 14 problems unaided. The AI group led early, and the pattern flipped on the final 3.
The author ties these results to a familiar pattern in learning research:
- Blocked practice beat interleaved practice during practice, but interleaving won at test.
- In a Make It Stick story, game-day curveball hitting improved only after interleaved training.
- Cramming helps on immediate tests, while spaced practice wins on delayed ones because it requires effortful retrieval from long-term memory.
- Rereading beats retrieval practice on immediate tests, but retrieval practice comes out ahead after a delay.
The author cautions that difficulty helps only within reason, since a task that's too hard can hinder learning. Education calls this Vygotsky's Zone of Proximal Development, and cognitive psychology calls it desirable difficulties. AI, the author argues, makes tasks far easier than earlier shortcuts, which raises the question of whether homework exists for grades or for learning. The post mentions the trend called friction-maxxing and closes with a Kurt Vonnegut passage about taking the slow way to mail his typed pages.