For decades, UX designers treated every extra second a user spent thinking as a failure. Don’t Make Me Think wasn’t just the title of Steve Krug’s book — it was the unspoken benchmark for our work. Then growth hacking in the 2010s turned removed friction into one of the go-to explanations for rapid product growth. Conversion optimization taught teams to see every additional step as a potential drop-off in the funnel. Mobile-first forced us to physically shorten scenarios.
The ones who began questioning this approach weren’t outside critics — they were practitioners. Cennydd Bowles, a former design lead at Twitter, argued back in 2019 that Krug’s principle doesn’t always hold. Sometimes it’s worth making users think by adding friction.
With the rise of AI, the need to challenge the frictionless ideal has become more pressing. A user flow can now lead not just from an already-formed intent to action, but from a machine recommendation to a decision the user makes with almost no scrutiny. The question of where to leave friction now will shape not just the quality of a single flow, but the kind of digital behavior we’re normalizing for years to come.
What Is Calibrated Friction in UX?
Calibrated friction is a design approach where intentional delays or extra steps are added to an interface proportional to the risk and severity of an action. Unlike frictionless design, it gives users the critical information and time needed to reassess high-stakes decisions before proceeding, preventing costly mistakes and passive acceptance of AI outputs.
Not More Friction, but the Right Friction
Good design should free people from work that doesn’t help them make a decision — but that doesn’t mean removing all effort. I don’t believe in frictionless design as a universal principle. Rather, I believe in calibrated friction, meaning that it’s proportional to the risk and consequences of an action. Where consequences are easy to reverse, an extra pause only complicates the scenario. Where an action is hard to undo or a mistake could prove costly, the interface should give the person a chance to reassess the outcome before proceeding.
That kind of pause also signals that the product respects the weight of the user’s action. A classic example is deleting a repository on GitHub: technically, a double-click could do the job, but instead you’re required to type the exact name of the repository. It’s slow and unpleasant, but it works as a signal. It tells the user, “We treat your repositories as something valuable.”
To determine what level of friction an action needs, ask yourself three questions:
What Happens If the User Makes a Mistake?
Two things matter here: whether the action can be reversed and how serious the consequences are. If a deleted file stays in the trash for 30 days, an elaborate confirmation is probably unnecessary — an undo notification is often enough. If the action is hard to reverse and its financial, legal, medical, reputational or security consequences are significant, friction before execution is justified.
How Often Does the User Perform This Action?
Repeated operational tasks don’t need friction. Adding it leads to alert fatigue and gets ignored. A rare action may warrant an extra check, but the deciding factor is still the severity of the consequences.
How far do the consequences reach?
The interface shouldn’t show an abstract “Are you sure?” but a concrete blast radius: who will be affected, what data will change and whether the previous state can be restored. Changing a setting in your own profile is not the same as applying it across an entire organization.
Calibrated friction doesn’t mean making an action difficult. Its job is to give the user the information they need to decide.
I saw this most clearly while working on a redesign of a B2B SaaS platform with multi-tenant settings. The problem was that admins regularly deleted configurations by accident, affecting dozens of client workspaces at once. This led to one or two incidents a week and support escalations.
We added an info banner at the entry to the section, indicating how many client workspaces a change would affect. By default, a preview mode opened where the admin could see exactly what would change and where. We also made Apply a separate, explicit step. Clicking it started a 30-second countdown, with a progress bar and a cancel button, before the change went live. It sounds excessive, but in practice, a notable share of actions get canceled in that window. Over three months, the number of incidents dropped significantly and the saved support time amounted to dozens of hours per month.
How Design Turns Trust in AI Into a Default
So far, we’ve been talking about interfaces that help a person carry out their own decisions. AI changes everything. Often, it proposes the decision itself. For instance, a system offers a suggestion the user can evaluate and reject. AI agents go further. The person doesn’t approve every step, but only sets the goal and the boundaries.
The problem is that interfaces for AI products have inherited the patterns product design spent decades refining to ease the path to action: defaults, a prominent primary CTA, one-click apply. But in a traditional interface, a single click shortened the path to a predictable outcome. In an AI scenario, it can shorten the path to real consequences and turn the evaluation of AI output into passive acceptance.
This effect is amplified by how the interface shapes perception of the system itself. AI often responds in a flat, authoritative tone, without flagging doubts or the limits of its own knowledge. The speed of a response is easily read as a sign of competence: The system answered instantly, so it must know. Sparkles, wand icons and phrases like “AI-powered” add an aura of magic, while a name, an avatar or a “Thinking…” indicator make the interaction feel like talking to a person. Users trust “Sam” more than “AI model X.”
At the same time, the interface often hides exactly the information that would give the user reason to doubt. The person sees a finished output but not the level of uncertainty behind it. This is how a trust illusion forms. It’s not because the user chose to fully trust the system, but because the interface didn’t give them enough opportunity to check whether that trust was warranted.
Are these conditions created deliberately? Yes and no.
Deliberately, it’s done to drive engagement or conversion. In many cases, “trust equals usage” is the business logic for AI startups looking to grow fast. A more honest UI would slow adoption. Unconsciously, it happens because of copying passive patterns from non-AI products where the stakes are lower.
And beneath both reasons lies the same metric as always. Acceptance rate rewards a removed step before an AI suggestion just as it once rewarded a removed step before an action.
What’s at Stake in a Frictionless World?
At first glance, the friction debate might look like an internal UX community dispute: whether an extra step is necessary, where to place a review, whether an additional confirmation will hurt conversion. But the stakes are far higher.
We are at a moment that resembles the early web in the 1990s or early social media in the 2000s. The patterns we create for AI products today will become industry standards. Think of infinite scroll. It originally solved a concrete UX problem — letting users browse content without constant interruptions. But combined with a business model that rewards time spent and engagement, the absence of a stopping point became a tool for capturing attention.
Infinite scroll has been around since 2006. In 2026, a regulator addressed these patterns for the first time: the EU labeled infinite scroll, autoplay and push notifications on TikTok — where these patterns are core to the experience — as elements of addictive design.
A locally successful UX pattern can become a standard before we see its long-term consequences.
AI is increasingly becoming a layer inside search, editors, B2B systems, financial products, medical products — the list goes on. The patterns forming around AI today won’t stay the language of a few chatbots. But when the decision about how much friction should precede an AI conclusion is calibrated against a metric, the choice will always lean toward less. And the consequences won’t be accidentally deleted files. They’ll be delegated judgment.
I believe the next three to five years are decisive. Either the industry finds an approach to honest AI design — one that shows limits, encourages critical thinking and applies calibrated friction to high-stakes AI actions — or we end up with the AI equivalent of the attention economy, with all its pathologies.
