UI UX Design
by OneZeroEight
Oct 05, 2026
6 mins read
For years, good UX had a simple enemy: friction.
Remove a click. Shorten the form. Skip the confirmation. Get the user from intent to outcome before they have time to wonder where the button went.
That logic still works for booking a cab or changing a profile photo.
But AI changes the equation.
An AI system doesn’t just wait for instructions. It can recommend, generate, infer and, increasingly, act. And when software starts making judgment calls, the fastest path isn’t always the safest or even the most useful.
For some AI products, friction isn’t a design flaw. It’s part of the product.
Here’s the thing: people have a habit of trusting machines when the machines look confident.
A 2025 global study led by the University of Melbourne with KPMG surveyed more than 48,000 people across 47 countries. It found that 66% of respondents relied on AI output without checking its accuracy, while 56% reported making mistakes in their work because of AI. That’s survey data, so it captures reported behaviour rather than experimentally observed errors. Still, the scale is hard to ignore.
Speed can make this worse.
Research published at CHI 2025 surveyed 319 knowledge workers across 936 real workplace examples. Researchers found that greater confidence in generative AI was associated with less critical-thinking effort. The work of thinking also changed: people moved from gathering information to verifying it, and from solving the task themselves to supervising AI output. Again, this was self-reported behaviour, not a controlled causal test.
That shift matters for AI UX design.
If your interface makes a probabilistic answer look like a deterministic one, smoothness can quietly become over-reliance.
The goal isn’t minimum friction. It’s minimum unnecessary friction.
There’s stronger experimental evidence here too.
In a controlled study of 199 participants, Harvard researchers tested “cognitive forcing” features that made people pause and think before relying on AI recommendations. These interventions significantly reduced over-reliance on incorrect AI advice compared with simpler explainable-AI interfaces.
There was a catch.
Participants liked the safer designs less. The interfaces that reduced over-reliance most received the least favourable subjective ratings.
That’s an awkward finding for product teams obsessed with satisfaction scores.
A little annoyance may sometimes mean the interface is doing its job.
Think about an AI drafting a campaign headline. Low consequence, easy to reverse. Let it fly.
Now imagine an AI agent sending that campaign to 300,000 customers. Same intelligence, very different UX problem.
The moment AI moves from “Here’s what I suggest” to “I’ve done it”, friction starts looking less like a nuisance and more like a seat belt.
This becomes especially clear with AI agents.
In its 2025 Operator system card, OpenAI described designing safeguards around both the potential harm of an action and how easy that action would be to reverse. Actions such as sending emails, deleting calendar events and completing purchases could require explicit human confirmation.
In internal evaluations, confirmations and related safeguards reduced the risk from model mistakes by roughly 90%. That figure comes from the developer’s own evaluation rather than an independent study, so it should be read in that context.
Still, the design principle is useful.
More autonomy should change the interface.
A chatbot suggesting a refund amount is one thing. An agent issuing the refund is another.
A useful AI product design rule is to think across four variables:
Consequence: What happens if the AI is wrong?
Uncertainty: How confident should we be in this output?
Autonomy: Is the AI suggesting something or executing it?
Reversibility: Can the user easily undo the result?
Low consequence + low autonomy + easy reversal? Keep the experience fast.
High consequence + uncertain output + autonomous action + poor reversibility? Add a checkpoint. Show the evidence. Preview the action. Ask for approval. Sometimes, don’t automate the final step at all.
This risk-sensitive approach also fits the broader direction of NIST’s Generative AI Risk Management Profile, which treats human oversight and the human-AI configuration as part of managing generative-AI risk rather than assuming automation itself is the goal.
But there’s another trap here.
More friction isn’t automatically safer.
A systematic review of 23 studies on clinical decision-support alerts found average override rates ranging from 46.2% to 96.2%. Many overrides were appropriate. In other words, bombard people with warnings and they learn to dismiss them.
A confirmation box nobody reads is theatre, not oversight.
This is where founders and product leaders may need to rethink the dashboard.
Time to completion, clicks, conversion and satisfaction still matter. But for AI products, they’re incomplete.
A five-second workflow isn’t successful if the user confidently approves the wrong answer.
Alongside speed, teams should measure appropriate reliance, correction rates, overrides, successful human interventions, reversibility and downstream outcomes.
That changes the product question.
Not: How quickly can we remove the human?
But: Where does human judgment create enough value to deserve a few extra seconds?
That’s a harder UX problem. It’s also a much more interesting one.
Frictionless UX isn’t dead. It just shouldn’t be a religion.
For AI products, the better goal is friction in proportion to risk.
What is friction in AI UX design?
Friction is any intentional step that slows or interrupts a user or AI system—such as confirmations, previews, source checks, warnings or human approval. Used well, it gives people a chance to inspect uncertain or consequential AI decisions.
When should an AI product add human approval?
Human approval becomes more useful as consequence, uncertainty, AI autonomy and irreversibility increase. Sending an email, publishing content, changing permissions or completing a transaction may deserve more oversight than generating a draft.
Can too much UX friction be harmful?
Yes. Repeated warnings can create habituation and alert fatigue. Users may begin clicking through prompts without thinking. The aim is targeted friction at meaningful decision points, not more pop-ups everywhere.
What should founders measure beyond AI task completion time?
Consider decision quality, appropriate reliance, error correction, human override rates, successful interventions and downstream outcomes. A fast AI workflow that produces unnoticed errors can look excellent on a traditional UX dashboard while failing users.
Is frictionless UX bad for AI products?
No. For low-risk, reversible tasks, removing friction can create a much better experience. The mistake is treating frictionlessness as a universal goal when AI can make uncertain decisions or take consequential actions.
Share
