Story
When an AI answer could lead me into a costly mistake, I want 1 plain warning about this answer, where I am reading, so I know what to check before I act.
Rosa, a nurse renting a flat in Chicago, asking about her deposit on her phone after a night shift.
- 01 Setting
Rosa asks if her landlord can keep her whole deposit for a small stain.
- 02 Trigger
The answer sounds sure. But the rules depend on where she lives.
- 03 Without it
The only warning is a grey "AI can make mistakes" under every answer. She has stopped seeing it.
- 04 With it
The answer ends with 1 line: general information, not legal advice, rules differ by city. And where to get help.
- 05 Outcome
She calls a free tenant help line before she writes to her landlord.
Live demo
This demo follows a script, so it plays the same every time - there is no real AI behind it. Change a setting to see the same moment play out differently.
What happens, step by step
| State | The person sees | The AI is | When | Screen reader |
|---|---|---|---|---|
| 1 | The question. | Starts. | 0 s | - |
| 2 | A pop-up before every answer. People tap I understand without reading. | Waits. | every time | “Dialog: AI can make mistakes” |
| 3 | The answer streams in (it arrives word by word). The quiet line under the composer (the box people type in) is always there. | Writes. | - | “Generating a response” |
| 4 | For high-stakes topics only: 1 line in amber at the end of the answer, about this answer, with a next step. | Writes. | end of the answer | - |
| 5 | The caveat stays with the answer, so it is there when people come back or share it. | Idle. | when done | “Response complete” |
On each device
- phone
- The caveat is part of the answer, so it scrolls with it. The quiet line under the composer may wrap to 2 lines - keep it short.
- tablet
- As on desktop.
- desktop
- The quiet line sits centred under the composer. The caveat sits at the end of the answer, with its link.
When to use it
- Answers can be wrong in ways that cost people money, health, or legal trouble.
- The answer depends on things the AI may not know - where people live, their contract, their history.
- People may act on the answer without checking anything else.
- People need to know which content the AI made, not what it may get wrong. Use AI label and disclosure instead.
- The AI cannot or should not answer at all - say so, and offer a way on. Use Error, refusal and limits instead.
- People need to see how sure the AI is about each part. Use Confidence display instead.
The parts of the pattern
- ACaveat in the answerAbout this answer and this topic, in 1 line, in amber. Only when the stakes are high.
- BNext stepWhere to check or get help. A warning with no next step only worries people.
- CQuiet line1 short, general line under the composer. Always there, never in the way.
Do and Avoid
Often not. A landlord can usually keep money only for…
⚠ General information, not legal advice. Rules differ by city.
Often not. A landlord can usually keep money only for…
AI can make mistakes.
⚠ General information, not legal advice.
Find free tenant help near you
Before you continue
AI can make mistakes. Our terms apply.
I understand
Writing and accessibility
- Say what may be wrong and why: "Deposit rules differ by city and state." Not only "AI can make mistakes".
- Give 1 next step: "Find free tenant help near you", "Check with your pharmacist".
- Keep the quiet line under 8 words: "AI can make mistakes. Check important info."
- Write it for people, not for lawyers. Put the legal text behind a link.
- Do not rely on amber alone. Start the caveat with a warning icon that has a text name, like "Caution".
- Keep the caveat inside the answer's text, so it is read in order with the answer.
- The quiet line must pass 4.5:1 contrast (the colour difference between text and background), even though it is quiet.
- Never trap keyboard focus (so people cannot tab out) in a pop-up that blocks every answer. If a 1-time notice is needed, people dismiss it once and never see it again.
Seen in real products
- Microsoft 365 Copilot ChatMicrosoft 365 app · Nov 2025The line "AI-generated content may be incorrect" under the compose box was hidden by default, "based on customer feedback". Admins can turn on a bolder version with a link to their own policy.Windows Latest, 2 Nov 2025 ↗
- SAP FioriDesign system · Nov 2026A general line - "AI can make mistakes. Check for accuracy." - goes below AI prompt inputs. Content-specific lines say "Generated with AI" or "Predicted with AI", then "Verify before use."SAP Fiori: AI Notice Base Concept ↗
- AWS CloudscapeDesign system · Nov 2026A disclaimer is an optional part of the chat, for legal or other important information. Teams decide with their product team whether they need one.Cloudscape: Generative AI chat ↗
Each example comes from the product’s help pages, release notes or teardowns, with the month it was checked. Products change every month - tell us when 1 is out of date.
Check yourself
- Apple HIG - Generative AI, raise awareness about hallucinations (updated Jun 2026) ↗
- Microsoft HAX Toolkit - G2 Make clear how well the system can do what it can do ↗
- Shape of AI - Caveat ↗
- Apple WWDC26 - Principles of great design ↗
- Apple WWDC25 - Explore prompt design & safety for on-device foundation models ↗
Seen this pattern done better, or a product that changed? Press Suggest next to any heading, or tell us about the whole page.