← All guides

AI + UX · 6 min read

Honest AI confidence & errors

Help people know when to trust AI — and how to recover when it’s wrong.

Explained like for kids

Sometimes a classmate sounds very sure but is totally wrong. Good AI UX is like a classmate who says “I’m pretty sure” or “I might be guessing” — and helps you fix mistakes fast.

Examples

Each example has a kid-friendly “why,” then a clear Don't and Do with a real UI mock so you can see the difference.

Guesses shouldn’t look final

Like for kids: Writing in permanent marker feels final. Pencil says “we can still change this.”

Don’t

Bold final copy for uncertain AI output

People accept hallucinations as truth.

Wrong — guess looks final

Your competitor’s churn is exactly 12.4%.

Stated as fact — may be hallucinated

Do

Draft styling + “Verify before sending”

Visual tone matches uncertainty. Invites review.

Right — draft + verify

Draft · verify before sharing

Estimated churn ~10–15% based on public reviews — confirm with your data.

Easy reject / undo

Like for kids: If the helper paints the wall the wrong color, you need a big eraser — not a scavenger hunt.

Don’t

AI change applied with no undo

Fear of trying AI. One bad edit ruins trust.

Wrong — no undo

AI rewrote your entire brief.

Done

Old version is gone

Do

Clear Undo + Keep side by side

Safe to experiment. Mistakes stay cheap.

Right — Undo / Keep

AI suggested a shorter brief.

UndoKeep

Human error messages

Like for kids: “Error 429 rate_limit_exceeded” is robot talk. “Too many requests — try again in a minute” is kid talk.

Don’t

Raw API / model error strings

Scary and useless for non-engineers.

Wrong — robot error
Error 429: rate_limit_exceeded at inference.gateway

Do

Plain language + what to do next

People recover without leaving the flow.

Right — human recovery

Too many requests right now

Wait about a minute, then try again.

Try again

High-stakes need a check

Like for kids: You wouldn’t take medicine because a cartoon character suggested it. Important answers need a grown-up check.

Don’t

One-click “Send invoice” from raw AI text

Financial / legal mistakes get expensive fast.

Wrong — one-click send from AI

Invoice total: $48,200 — ready to send.

Send invoice now

Do

Review checklist before the final action

Human-in-the-loop for irreversible steps.

Right — review checklist
  • ☐ Amount matches quote
  • ☐ Client email correct
  • ☐ Tax line reviewed
Send after review

Remember these

  • Don’t dress guesses as certain facts.
  • Offer “check this” paths for high-stakes answers.
  • Make undo / reject easy when AI is wrong.
  • Explain errors in human language, not model codes.

Trust and hallucination patterns for AI product interfaces. Rewritten in plain language for learning — not a reprint of the originals.