The Case for Friction in Human-AI Systems
Image generated by fal-ai/nano-banana-2, 2026.
Effort Is a Signal
Many AI products treat friction as the enemy: fewer clicks, faster drafts, instant answers. But in learning and decision-making, effort is not just a cost. It is also a measurement. The attempt before the answer reveals what a person understands, where confidence is misplaced, and whether help is filling a gap or merely masking it. Without that comparison between expected and actual performance, self-calibration has little to update.
Recent preprints on AI-assisted learning (arXiv:2605.21629) and reflective decision support in HCI (arXiv:2605.22509) both point in this direction. Their shared implication is that systems optimized for immediate task completion can improve short-term output while weakening transfer, error detection, or independent reasoning. That fits a long line of education research: Robert and Elizabeth Bjork’s work on desirable difficulties and Manu Kapur’s studies of productive failure both show that carefully structured struggle can improve long-term retention and transfer. In other words, less effort is not always better augmentation.
Friction as a Control Surface
“Every good regulator of a system must be a model of that system.” — Conant and Ashby
From a cybernetic perspective, friction is damping. A human-AI system that only optimizes speed invites overshoot: automation bias, premature acceptance, and brittle dependence. If the model decomposes the problem, writes the first draft, evaluates the result, and hides its uncertainty, the human loses observability of both the task and the self. The loop becomes an open-loop handoff.
Well-designed friction restores feedback. It might ask a user to predict an answer before reveal, rate confidence before seeing model advice, compare two options, or inspect provenance and uncertainty before accepting a recommendation. In education, this is why tutors have long used stepwise hints and faded scaffolding rather than instant solutions. An LLM tutor that asks a student to explain the next algebraic move may feel slower than one that prints the derivation, but it preserves the generation effect—the learning benefit of producing an answer before seeing one—and exposes misconceptions. In professional decision support, ordering matters too: clinicians who make an initial assessment before seeing AI advice can be better at spotting bad suggestions than those who see the model first and anchor on it. Even GPS navigation offers a mundane version of the same tradeoff: turn-by-turn guidance gets users there efficiently, but route previews and choice points do more for spatial learning.
Designing for Legible Delegation
Bad friction is bureaucracy; good friction is scaffolding. The best version is adaptive: it slows novices more than experts, and slows everyone more when uncertainty or consequence rises. High-stakes decisions, early skill development, low model confidence, and anomalous cases are the moments when an interface should create a pause. Friction should also make delegation legible. If the AI drafted the paragraph, selected the evidence, or classified the image, the interface should say so—and show how. Provenance markers, uncertainty bands, and “show work” views turn opaque outsourcing into inspectable cooperation.
That design stance echoes Bainbridge’s classic “ironies of automation”: remove too much routine effort, and people lose the practice needed when the system fails. The next benchmark for human-AI systems may therefore be not just time saved but judgment preserved. If the best assistant is not always the fastest one, what would change if AI were evaluated by the capability it leaves behind, not merely the effort it removes?