Independent briefing Learning & education
Choosing to stay human
Why friction, effort, and productive struggle still matter when a model can produce the answer instantly.
Independent briefing of One Useful Thing (2025). Read the original for its full evidence, context, and wording.
Start here
Three questions this article answers
Can an AI answer improve homework while weakening learning?
Yes. Students can submit better work with unrestricted assistance yet perform worse later without it.
What is productive struggle?
It is the effort of retrieving, testing, and revising ideas—the cognitive work that helps knowledge become durable.
Can AI tutoring improve learning instead?
Yes, when the system guides rather than substitutes and is designed around the learner’s next step.
Go deeper
Three related questions it doesn’t answer—matched to established papers
What happens when students receive unrestricted generative-AI help?
A randomized study separates short-term assisted performance from later unassisted learning.
Paper answering this questionGenerative AI Without Guardrails Can Harm LearningCan a well-designed AI tutor outperform classroom active learning?
A randomized controlled trial compares an AI tutor with an active-learning class.
Paper answering this questionAI Tutoring Outperforms In-Class Active LearningCan AI help human tutors teach more effectively?
A large randomized trial studies real-time AI guidance for tutors.
Paper answering this questionTutor CoPilotThe useful summary
What to take away
The piece separates performance from learning. A polished assignment is evidence that a product was produced, not that the student can explain or reproduce the underlying thinking. When AI removes retrieval, false starts, and revision, it can remove the mechanism by which a learner builds skill.
The alternative is not an AI-free classroom. It is AI that behaves more like a good tutor: asking for an attempt, diagnosing a misconception, offering a hint, and slowly withdrawing support. The design question becomes “what should the learner think through next?” rather than “how quickly can the system finish?”
Final conclusion
The three answers, brought together
AI can improve visible homework while harming durable learning because productive struggle is part of learning, not wasted time. Tutoring systems can help when they preserve that struggle and adapt support instead of replacing the student’s thinking.