Wrong is a normal state
Language models can produce fluent text that is simply incorrect. Designers cannot remove that risk. They can decide how the product behaves when it happens.
Nielsen's classic heuristics still apply. Help users recognize, diagnose, and recover from errors. Keep them informed about system status. Give them control and freedom to back out. AI does not retire these rules. It makes them more important.
Match guardrails to the stakes
| Stakes | Example | Design response |
|---|---|---|
| Low | A suggested email subject | Show it, let the user edit, no extra friction |
| Medium | A summary of a customer call | Link to the source, flag unverified claims |
| High | A change to billing or permissions | Preview the effect, require approval, log it |
| Critical | Medical, legal, or financial advice | Keep a human expert in the loop, restrict scope |
Say "I am not sure" like a person would
A percentage like 72% means little to most people. A sentence like "I could not find this in your documents, so this is a guess" tells them exactly what to do. Use plain flags where uncertainty changes the next step, and stay quiet where it does not.
Make correction one click
- Put a visible way to correct or flag an output next to it, not in a menu.
- Offer reasons as quick choices: wrong fact, wrong tone, not relevant.
- Let users retry with the correction applied, and confirm it worked.
- Feed corrections back to your team so recurring failures get fixed in the product.
Empty and failure states matter too
When the model cannot answer, do not show a generic error. Explain what happened, what the user can try, and offer a non-AI path to finish the task. A dead end is the worst failure mode for a feature meant to save time.



