The Value of Friction

The exam after the help

In 2025, researchers gave nearly a thousand secondary-school maths students access to GPT-4 while they practised.¹ One group used a standard chat interface. Another used a tutor version designed to give hints instead of answers.

During practice, both groups flew. Grades rose by 48% with the standard interface and by 127% with the tutor. Then the AI was taken away for the exam. Students who had used the standard interface scored 17% lower than students who had never had access at all. The tutor group largely avoided that drop.

The help worked. That was the problem. It removed exactly the effort that would have turned practice into skill.

Ease of performance and growth of capability are different things to optimise for.

Friction is not good

I want to be careful here, because this argument is easy to misuse. Friction is not a virtue. Forms that ask for the same information three times, approvals nobody reads, security steps that lock out people with a disability — that is waste, and removing it is a form of respect.

Much of the work I do in technology is about removing exactly that kind of friction, and I believe in it.

But “remove friction” has quietly become a universal design principle, and it isn't one. Some friction is doing work. The question is not whether something is effortful. It is what the effort protects.

The friction worth keeping

Protective friction sits between an impulse and an irreversible act: the confirmation before a payment, the few seconds to undo a message, the human review of a decision that changes someone's life. It protects safety and consent.

Developmental friction is the difficulty that builds capability. Trying to recall something before re-reading it improves long-term memory more than reading it again.² Attempting a problem before asking for help is how judgement forms. That is what the maths students lost.

Relational friction is the effort inside understanding another person: disagreement, explanation, apology, repair. If we automate the difficult conversation away, we often automate away the relationship it was building.

Constitutive friction is different in kind. It does not merely make an activity harder; it is part of what makes the activity what it is. The rules of chess are constraints, but remove them and you have not made chess frictionless — you have stopped playing chess. A sonnet's form is not an obstacle to the poem. A fast's restrictions are not usability defects in fasting.³ A difficult conversation cannot always be optimised into an automated message without changing the nature of what one person is doing for another.

That gives a sharper test than “the effort is the point”:

If removing the friction changes not just how easily something is done, but what it is, the friction may be constitutive.

What AI changes

Generative AI is the most powerful friction remover we have built. That is mostly good news. For many tasks, I want the answer, not the struggle.

But it forces a design choice that used to be invisible: what is this person here to do? If the purpose is to complete a task, remove the difficulty. If the purpose is to learn, to judge or to decide something consequential, the product should protect the part of the effort that builds the capability — hints before answers, drafts before sending, a human before a final decision.

The lesson from the maths study is not “make AI difficult”. It is: preserve the part of the difficulty that produces what the person came for.

I feel this most clearly in my own writing — including this essay, which I am developing with AI. When I write, I often know the territory before I know the argument. The frustrating part — moving paragraphs, deleting sentences, discovering that the point I thought I was making isn't quite the point — can feel inefficient. AI can remove much of that friction. But if I ask it to resolve the argument too early, I may get better prose before I have done the thinking that makes the prose mine.

For me, the useful boundary is not between writing with AI and writing without it. It is between using a tool to support the work and using it to bypass the part of the work that changes my own understanding. If AI formats a citation, I have lost nothing. If it settles what I believe before I have worked it out, something categorically different has happened.

Who carries the friction?

Friction is never evenly spread. A security step that protects an organisation can be pure burden for the user who has to repeat it every day. A synchronous meeting that feels rich to one colleague can exclude another because of time zones, caregiving or language. An extra confirmation that prevents one mistake can make a tool unusable for someone with a motor impairment.

So whether friction is protective or wasteful depends partly on whose perspective you measure. Good design asks that question explicitly.

Four questions before removing it

1. What value does this friction serve — safety, consent, learning, trust, something constitutive, or nothing?
2. Who bears the cost, and who receives the benefit?
3. Who controls it, and can it be bypassed in a real emergency without destroying it?
4. What would happen if it disappeared?

If the answer to the first question is “nothing”, remove it without regret. If it protects something real and the burden is fair, keep it — on purpose, and with evidence.

The right friction

We have spent decades treating every pause, every effort and every limit as a defect waiting for a fix. Many were. But some of those limits were where learning happened, where consent was given, where a relationship was tested.

And some were more than useful. The study at the start of this essay shows that friction can produce capability. My own writing tells me something further: sometimes the friction does not just lead to a better result. It is part of what makes the work mine.

The goal of good design is not minimum friction. It is the right friction, in the right place, for the right reason.

Sources

The full evidence base is on the Human Limits research page.

1. Bastani, H. et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26). doi:10.1073/pnas.2422633122. ↩
2. Roediger, H. L. & Karpicke, J. D. (2006). Test-enhanced learning. Psychological Science, 17(3), 249–255. doi:10.1111/j.1467-9280.2006.01693.x. ↩
3. Qur'an 2:183–187. quran.com. ↩