Skip to content
Ben Serkan Zararsız

Design for the moment AI is wrong

2025 was the year of AI fatigue. 2026 is the year trust became a design problem. Six patterns that make an AI feature feel reliable, especially when it is not.

On this page
  1. Intro
  2. Suggest, don’t overwrite
  3. Say what it is
  4. Show doubt where it matters
  5. Fail like a good colleague
  6. Same behaviour, every screen

Every product I worked on this year had an AI feature. Most of them had the same problem: they worked eight times out of ten, and nobody had designed the other two. The suggestion that was confidently wrong. The summary that invented a date. The “improved” text that overwrote the original. One bad moment, and people stop using the feature for good.

After a year of sparkle icons, it is clear what builds confidence: transparency, control, consistency, and help when the system fails. All four are interface work, and these are the six patterns I now look for in every AI feature.

Suggest, don’t overwrite

The most important decision is who acts. An AI that proposes keeps the person in charge. An AI that applies takes something away, and people notice. A suggestion lives next to the original, can be compared, edited, dismissed and applied with one tap. An automatic change needs an undo, every time, and even then it leaves the taste of lost control.

The platform guidance is blunt: let people dismiss what they do not want, revert or retry what they disagree with, and show where AI was involved. Keep the manual way within reach. If the feature disappeared tomorrow, the task must still be possible.

Reply draft

Thanks for your message, we will get back to you as soon as possible regarding your request about the invoice.

Suggested by AI

Thanks for your message. We’ll get back to you about the invoice by tomorrow.

Applied ·

A suggestion next to the original, apply or dismiss

Reply draft ✨ Improved by AI

Thanks for your message, we will get back to you as soon as possible regarding your request about the invoice.

Overwritten on the spot, no way back

Say what it is

People are not fooled for long. Content produced by a model should say so, where it appears, in plain words. In a survey this autumn, 60 percent of shoppers said that knowing a product image was AI-generated would lower their confidence to buy. Hiding it only moves the moment of distrust to later. Think of the label as expectation management.

The same goes for limits. A short line about what the feature can and cannot do, shown when it is introduced and again at the moment an error is likely, beats a permanent “may make mistakes” footer that everyone has stopped reading.

AI summary · from 3 reviews

Runs small, order half a size up. Waterproof held up in heavy rain. The laces wear out early.

Labelled, with the sources it came from

Runs small, order half a size up. Waterproof held up in heavy rain. The laces wear out early.

Same text, dressed up as product copy

Show doubt where it matters

A model’s confidence score is no real measurement. Uncertainty can still be designed. In anything that costs money or time, show it: a “please check” marker on the one field the system was unsure about, a first-person “I could not read the date”, a range instead of a precise number. An assistant that admits what it does not know earns more trust, not less.

Sources help only if they sit next to the claim they support and link to the exact passage. A row of citations at the bottom makes an answer look more reliable than it is, and almost nobody clicks.

Read from receipt

VendorCafé Kittel
Amount€18.40
Date14 Nov 2026?Hard to read, please check
CategoryMeals

One field flagged, the rest confident

Read from receipt

VendorCafé Kittel
Amount€18.40
Date11 Nov 2026
CategoryMeals

Everything confident, one value wrong

Fail like a good colleague

The moment AI is wrong is the moment your interface is judged. “Something went wrong” is not a design. A good failure says what happened, keeps what the person typed, offers the manual path and does not blame them. The same applies to a slow answer: show progress, allow cancelling, never lock the screen.

And remember the task itself. In a study of an agent that orders food, the problems were not model failures but basics: fees shown only at checkout, an address revealed without explanation, a carousel that hid how many options there were. Efficiency does not build trust. Information does.

Expense

I couldn’t read the amount on this receipt. You can type it, or try another photo.

Says what happened, offers the manual way

Expense
Something went wrong. Please try again later.

Generic error, everything disabled

Same behaviour, every screen

Trust grows from repetition. The same entry point, the same tone, the same way to accept or reject, the same place for the label. A feature that behaves differently on every screen feels random, and random is what people fear most in a system that already guesses.

In my projects, an AI feature only ships when these points hold:

  1. The person decides. AI suggests; it never applies without an undo.
  2. AI content is labelled where it appears.
  3. Uncertainty is visible where it matters, and sources sit next to claims.
  4. Failures keep the input and offer the manual way.
  5. The feature is optional, and the task works without it.
  6. Same behaviour, every screen.

Nobody remembers the sparkle icon. People remember how the feature behaved when it got something wrong.

Sources

  1. Apple Human Interface Guidelines: Generative AI
  2. Baymard Institute: AI Engagement: Quantitative UX Insights (October 2026)
  3. Nielsen Norman Group: Explainable AI in Chat Interfaces (2025)
  4. Nielsen Norman Group: Designing AI Agents: 4 Lessons from China’s Qwen Agent (2026)