Interfaces that think for us. The drift is decided by the UX team

I've spent years writing with a machine that closes my sentences before I do. The unsettling part isn't that it gets them wrong. It's that it gets them right. Don Norman, in The Design of Everyday Things (Basic Books, 1988; revised ed. 2013), called friction the gap between what the user wants and what the system hands back, and advised narrowing it with clear signifiers, visible affordances and feedback. Good advice for a tool that performs a specific task. Carried over to the interfaces that mediate our own thinking, it produces an effect nobody in 1988 had any reason to foresee: when the friction is erased, so is the gap in which the user was thinking. And the direction of that erasure isn't set by us.

Friction as cognitive space

The instant between intention and action isn't empty. Things happen inside it. You decide how you're going to say what you want to say, you choose between two ways of framing it, you retrieve the fact you needed, you hesitate for a second, you correct yourself. Friction is a nuisance, and that's why reducing it improves the feel of use. But that same nuisance is the material condition of operations we take for granted: working memory, the search for the right word, the judgement about whether what you've written says what you meant.

When the system offers you the next word, that instant shrinks. The suggestion is statistically likely, it's usually correct and it almost always turns out acceptable. The problem lives precisely there, in how reasonable the proposal is: accepting it costs less than thinking up your own. That the path of least effort wins when everything else stays equal is a commonplace of cognitive psychology, and it's best not to pin it on a specific experiment or dress it up as a law. The statement is simple and the measurement is fiendish: your own phrasing gets exercised less.

What has vanished in a decade

It's worth naming the frictions that contemporary design has eliminated on purpose, because in isolation each one looks harmless. Aggressive autocorrect amends without asking, often without the user even noticing the change. Predictive autocomplete closes the sentence before you decide how you wanted to close it. Suggested replies —Gmail, Outlook, WhatsApp are right there— hand you the whole message already drafted, ready to send. The search box proposes what to search while you're still framing the query. The assistant that regenerates invites you to hit the button again instead of rephrasing what you asked. And the automatic summary serves the document pre-chewed, so you end up judging the summary and not the text.

Each of those pieces is useful in its place and answers a sensible technical decision. What changes is the sum. Multiplied by thousands of interactions a day over years, that collection of small comforts reorganises the relationship we have with information and with language.

The MIT study and the symptom

In June 2025 the MIT Media Lab put out a paper titled Your Brain on ChatGPT. Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task (preprint, arXiv:2506.08872). Fifty-four participants wrote essays with electroencephalography on, split across three conditions: with a language model, with a search engine and with no help at all. The authors report lower connectivity and lower activation in the regions associated with executive function when writing with the model, compared with writing unassisted.

Taking it with tongs isn't rhetorical caution, it's what the work asks for. A preprint without peer review, a small sample, a single tool, and a methodology its own authors warn shouldn't be extrapolated. None of that authorises talk of brain damage, and I'll say so before someone says it for me.

What the study does do well is frame the question. That a tool reduces the effort of doing what it was designed for surprises nobody; that's its definition. The question that matters is the other one: what exact cognitive task are we delegating, and what happens to that task when it stops being performed, repeatedly and silently, over years.

Buçinca and the cost of trusting without thinking

Zana Buçinca, Maja Malaya and Krzysztof Gajos (CSCW 2021), in To Trust or to Think. Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-Assisted Decision-Making, set up two experiments —finding shortest paths in graphs and fixing spelling errors— and found that users tend to accept the AI's suggestion even when it's wrong and they have the information needed to spot the failure.

Their most interesting finding carries an unfortunate name: cognitive forcing functions. They're frictions placed on purpose —requiring a justification of the decision, delaying the system's response, demanding an explanation— that cut overconfidence and improve the quality of judgement along the way.

From which follows something the design industry would rather not hear. If you want the user to judge well what the machine proposes, you have to give back part of the friction you took away. Practice walks in the opposite direction. And it walks happy.

The metric that decides

The reason for that opposite direction is no mystery. The metric that rules a product is retention: time inside the app, sessions per day, daily active users. Retention rises when perceived friction falls. More retention justifies more development resources, which produce even smoother interfaces, which improve retention again. The cycle feeds itself, and that's why it's so hard to break from inside.

The metrics that could counterbalance it get measured little, and not out of carelessness. That the task came out well done, that the decision was correct, that the user learned something, that they're still satisfied a month later: all of that takes longer to give a signal, costs more to turn into a number and doesn't touch revenue with the same immediacy. You don't need bad faith to prefer the other thing. A calendar that closes every three months is enough.

Arunesh Mathur and his co-authors (CSCW 2019), in Dark Patterns at Scale. Findings from a Crawl of 11K Shopping Websites, crawled around 53,000 product pages across some 11,000 stores and documented 1,818 instances of patterns designed to push the user towards decisions that benefit the service and not them. The dark pattern, the openly manipulative design, is just the visible extreme of something rather more widespread: the mismatch between what the product measures and what suits whoever uses it.

The double effect

Erasing friction produces two things at once, and it's worth not mixing them up. On one side it improves performance on the task in front of you: if what you need is to get an email out, the suggested reply gets it out, and if you write repetitive code, autocomplete raises your lines per hour. On the other side it slowly wears down the capacity that friction was exercising. What you systematically delegate —shaping an idea, weighing an option, building an argument from scratch— gets worked less, and what gets worked less comes out worse the day you have to do it without a net.

Both effects are real. The conflict isn't in the first one, but in how they're distributed over time. The benefit arrives today; the atrophy arrives late. Design decisions close on a quarterly horizon while the cognitive consequences take years to show their heads. That asymmetry isn't innocent: it pushes, structurally, time after time, towards the same place.

Norman, the signifiers and a formula that lost its context

Norman recommended reducing friction against a very specific evil: interfaces that left the user lost with ambiguous signifiers, unclear affordances, nonexistent feedback. His world was that of the physical object —doors you don't know whether to push or pull, controls that don't say what they do— and that of the productivity software of the eighties and nineties. In that world, reducing friction meant clarifying.

The slogan survived its world; the method changed underneath. The imperative was kept intact —reduce the friction— and its content swapped: no longer clarify so the user decides better, but predict and decide in their place. The interface that clarifies lets you decide. The one that decides spares you the trial of deciding. Both are advertised with the same slogan and produce opposite cognitive effects, and confusing them isn't a stray slip: it's the operation industrial practice has been normalising for years. Steve Krug, in Don't Make Me Think, Revisited (New Riders, 2014), took the slogan to its purest version, that of treating every question surfacing in the user's head as a cost to be erased. It works wonderfully for a checkout process. For a thinking tool, unqualified, it doesn't.

The different cognition

Out of this doesn't come, necessarily, a worse cognition. Out of it comes a cognition with a different profile. Faster to respond and clumsier when building alone, more solvent in what it already knows and more lost when something new appears, more comfortable with an assistant alongside and more insecure the day it doesn't have one. The difference is operational, not metaphysical, and that's precisely why it's so hard to see.

The serious part is that this profile wasn't chosen as anyone's goal. No education ministry set it, no health committee, no civic debate. It's fallen out as a by-product of design decisions oriented to retention. The generations growing up with an assistant from childhood will have a cognition different from those who grew up without one. The difference won't be minor. And nobody is measuring, systematically, what's gained and what's left behind along the way.

The mandate nobody audited

A political question remains that almost never gets put on the table. The design teams of the handful of companies controlling the digital cognitive infrastructure decide, month by month, where the thinking of hundreds of millions of people shifts. That mandate was granted by no one. No authority audits the effect of those decisions on aggregate cognitive capacity, nor is there any obligation to measure it, to report it, or to justify it. The sector self-regulates with metrics that, by their very construction, are blind to that effect.

What comes now is my opinion, though I'll defend it. The distance between the size of that mandate and the total absence of oversight is hard to match in recent history. Radio, television and the school ended up, sooner or later, subject to rules about what they broadcast and how, precisely because their effect on the collective mind was recognised as a public matter. The interface that this afternoon finished three of our sentences hasn't yet reached that conversation. And in the meantime, it keeps finishing them.

Definitions

Friction (in UX) is the effort, cognitive or physical, the user has to put between their intention and the result when handling a system. Reducing it improves speed and comfort, and tends to increase use.

Signifier is the perceptible indicator that communicates where and how to act. It's a central concept in Norman; the English term is rendered here.

Affordance is the action an object offers whoever uses it. The term comes from Gibson and Norman popularised it.

Cognitive forcing function is a design that introduces friction on purpose to force reflection before deciding. The concept is from Buçinca and her co-authors.

Retention (engagement) groups the usage-time metrics —time in the app, sessions per day, daily active users— dominant in the interface-design industry.

Dark pattern is a design deliberately oriented to manipulate the user towards behaviours that benefit the product and not them.

References

Norman, D. The Design of Everyday Things. Basic Books, 1988; revised ed. 2013. Origin of the concept of friction and of the signifiers and affordances that structure the article.

Buçinca, Z., Malaya, M. & Gajos, K. "To Trust or to Think. Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-Assisted Decision-Making". Proceedings of the ACM on Human-Computer Interaction (CSCW1), 2021. Basis of the section on overconfidence and cognitive forcing functions. arXiv:2102.09692.

Mathur, A., Acar, G., Friedman, M., Lucherini, E., Mayer, J., Chetty, M. & Narayanan, A. "Dark Patterns at Scale. Findings from a Crawl of 11K Shopping Websites". Proceedings of the ACM on Human-Computer Interaction (CSCW), 2019. Source of the dark-pattern count and the mismatch between product metric and user. arXiv:1907.07032.

MIT Media Lab. Your Brain on ChatGPT. Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. Preprint, June 2025, arXiv:2506.08872. EEG study with 54 participants cited in the section on the symptom; preliminary and not peer-reviewed.

Krug, S. Don't Make Me Think, Revisited. New Riders, 2014. The extreme version of the reduce-friction slogan discussed in the article.

Carr, N. The Glass Cage. Automation and Us. W. W. Norton, 2014. General framework on the effects of automation on human capacity.

Further reading

Sherry Turkle. Reclaiming Conversation. The Power of Talk in a Digital Age. Penguin Press, 2015. Her argument on the technological mediation of conversation carries over easily from the social to the cognitive.

Maryanne Wolf. Reader, Come Home. The Reading Brain in a Digital World. HarperCollins, 2018. Documents the change in sustained reading under digital environments; the framework applies to the change in linguistic phrasing under generative assistance.

Jonathan Haidt. The Anxious Generation. Penguin Press, 2024. A methodological precedent on how retention-oriented design produced massive unanticipated effects, useful before dismissing analogous concerns about conversational AI.

You may also be interested in

• UX as gentle manipulation
• The design of dependence
• Behaviour-oriented design
• What happens to your brain when you use ChatGPT, according to MIT

Elsewhere

• A Review of the Negative Effects of Digital Technology on Cognition (arXiv)
• Supporting Cognition With Modern Technology: Distributed Cognition Today and in an AI-Enhanced Future (PMC)