Generative AI can make almost any task feel effortless. Draft an essay, design an experiment, summarise a chapter — all in seconds. But a new paper by Emily Zohar, Paul Bloom, and Michael Inzlicht, published in Communications Psychology, argues that this smoothness is not a pure gain. Removing friction may quietly remove the conditions under which real learning happens.
The authors’ central claim is that friction — the difficulty, effort, and even frustration involved in doing hard cognitive work — is not an obstacle to learning. It is the mechanism through which learning occurs. Struggling to retrieve, reorganise, and apply information build durable understanding. When AI tools remove that struggle entirely, they don’t just save time; they remove the very process that would have produced memory, skill, and a sense of ownership over the work.
[Productive friction increases learning, meaning & motivation. This is akin to the visual metaphor on the right. Taking a chairlift and hiking can both get us to our destination but the latter method is more effortful and helps to build us physically.]
Zohar and colleagues draw on the established research base around desirable difficulties — decades of cognitive science showing that effortful retrieval and problem-solving produce stronger, longer-lasting learning than passive or frictionless study methods. Building on this, they cite evidence that heavy AI users recall their own completed work less accurately afterward, and perform worse once the AI is no longer available to assist them. The work looks finished. The learning behind it often isn’t.
Importantly, the paper does not argue that more struggle is always better. The relationship between friction and learning benefit follows an inverted U-shape. A moderate amount of difficulty strengthens learning. Too little friction, and there’s nothing for the brain to work with. Too much, and the learner is overwhelmed rather than engaged. The goal isn’t maximum difficulty — it’s landing in the productive middle, the zone educators have long called productive struggle.
This has direct implications for how AI gets used in teaching. If students are simply handed a finished answer, the surface output improves while the underlying learning quietly declines — a pattern that closely echoes findings from other recent studies on AI use in schools. The task for educators, then, isn’t to eliminate AI, but to deliberately reintroduce effort at the points where it matters most.
Let students brainstorm with AI, but require them to defend what they kept and what they discarded. Sequence the work so students solve first, then check with AI — never the reverse. Treat AI-assisted output as a starting draft to be interrogated, not a finished product to be submitted.
The friction we remove for a student today is friction they never learn how to handle tomorrow. This paper is a useful reminder that in education, ease and learning are not the same thing — and sometimes they pull in opposite directions.
