Back to News
UX writing AI editing plain language

How to Humanize AI-Written Error Messages Without Hiding the Problem

October 6, 2026

Making AI-written copy sound human is not always about adding warmth. In an error message, the most human thing you can do is explain the problem and offer a useful next step. A cheerful sentence that hides the cause of a failed action is still bad writing.

This guide focuses on one small but important editing task: rewriting AI-generated form errors and support microcopy without losing the facts a user needs. It is not a method for concealing AI use or bypassing a detector.

Start with the actual failure, not the tone

Before rewriting, ask the product owner what happened. Did a required field remain empty? Did an upload exceed the supported size? Did the service fail after accepting a request? Those are different events and should not receive the same friendly apology.

The GOV.UK Design System's error-message guidance offers a useful foundation: explain what went wrong and how to fix it. It also distinguishes validation errors from problems users cannot fix, such as a service-side outage. That distinction should survive every AI-assisted rewrite.

Do not invent a cause just to make the text more specific. If the application cannot determine why a payment failed, the copy should not assert that a card has expired. Better wording cannot compensate for missing diagnostic information.

Use a three-part editing brief

Give your writing tool three inputs: the verified condition, the action the user can take and any wording that must remain unchanged. For example:

Rewrite this upload error in clear, calm English. Verified condition: the file exceeds the product's configured limit. Required next action: choose a smaller file. Keep the supported maximum exactly as supplied by the application. Do not invent file types, limits, contact details or guarantees. Return one short message.

Keep real customer details, account numbers and confidential support transcripts out of the prompt unless your organization has approved that data use. A synthetic example usually provides enough context for a wording exercise.

Before and after: five illustrative rewrites

These examples are editorial demonstrations, not messages from a tested product. Replace every bracketed value with a verified application value before use.

  • Vague: “Oops! Something isn't quite right.” Clearer: “Enter your email address.” Use this only when the address field is empty.
  • Technical: “Payload exceeds permitted threshold.” Clearer: “Choose a file smaller than [verified maximum size].” Check whether the product's limit is strictly smaller than or at most the displayed value.
  • Blaming: “You entered an invalid date.” Clearer: “Enter the date in the format [required format].” Use only if format is the actual problem; a validly formatted future date needs a different explanation.
  • Overconfident: “Don't worry, your payment definitely wasn't taken.” Clearer: “We could not confirm the payment. Check your order status before trying again.” The final wording must reflect the actual payment workflow and available status page.
  • Unhelpfully friendly: “Our servers are having a little nap!” Clearer: “The service is temporarily unavailable. Try again later.” Add a status-page link only if one exists and is relevant.

Notice that the best version is not always shorter. Sometimes a few extra words prevent a duplicate action or explain a constraint. Brevity is useful when it removes clutter, not when it removes essential information.

Keep the user's task visible

A good error message should let the reader answer: which item needs attention, what needs to change and what happens next? If the same message appears beside several fields, include enough context to distinguish them.

Avoid “click here” as the only link text. Describe the destination or action, such as “Check order status.” Keep button labels aligned with what the interface actually does. “Continue” and “Submit payment” create different expectations.

For forms, wording is only part of accessibility. The GOV.UK guidance recommends showing validation errors next to their fields and in an error summary. Your implementation also needs appropriate associations and focus behavior. A clear sentence alone is not a claim that an entire interface meets accessibility requirements.

Review meaning separately from voice

Use two passes. In the first, compare the rewrite with the verified product behavior. Check limits, dates, required fields, permissions and promises about saved work. In the second, edit rhythm and tone. Remove excessive apologies, jargon, jokes that trivialize a problem and vague reassurance.

A content difference checker can help you notice changed words, but a diff cannot decide whether a changed claim is true. Have someone who understands the feature review the final message in context.

If you use HumanizeAIText to explore a more natural version, treat the result as a candidate, not an approved interface string. Detector scores do not establish usability, factual accuracy or accessibility. Test whether people can recover from the error instead of optimizing for a score.

A release checklist for rewritten microcopy

  • The failure condition and suggested action are accurate.
  • Product limits and variable placeholders remain intact.
  • The message does not blame the user for a service problem.
  • No unsupported promise is made about payments, saved work or response times.
  • Links and buttons lead to the stated destination or action.
  • The message makes sense alongside its field and in the wider page.
  • A reviewer has tested both the error and recovery path.

Natural writing is useful when it makes the reader's next decision easier. For error messages, clarity and honesty are the voice worth protecting.

Source: GOV.UK Design System: Error message, accessed October 6, 2026. The source provides design guidance; the rewrite examples and review process here are editorial adaptations, not measured performance claims.