Marcus Andrey Study
Ler em português Library

AI for UX · Step 2 · 7 min read

AI for content design

How to use AI for UX writing without losing your product's voice: what it does well and badly, how to give it your voice and terminology, how to review AI copy, translation versus localisation, inclusive language, and when to tell users that text was written by AI.

Writing is the most common way UX practitioners use AI. In an NN/g survey of 841 practitioners, more than three quarters used it for text: microcopy, error messages, labels that must fit a character limit, changes of tone (NN/g). It's fast and often good. It's also where AI quietly drifts away from your voice, your terms and your facts. This note shows how to get the speed without the drift.

What you'll be able to do after reading:

  • Know which writing tasks AI does well and which it does badly
  • Give AI your voice, tone and terminology so it writes like your product
  • Review AI-written copy with a short, repeatable checklist
  • Tell translation from localisation, and know when a person must review
  • Decide when users should be told that text was written by AI

1. What AI is good and bad at in UX writing

Good at Bad at
Many variations quickly ("10 ways to say this") Knowing your product, users and rules
Fitting a character limit Staying in your voice over many screens
Changing tone (more formal, friendlier, shorter) Getting facts right: prices, limits, policy
Consistency checks against a list of terms Knowing when not to say something
First drafts of help articles and empty states Cultural nuance in other markets

Practitioners in the NN/g survey reported the same risks: output too formal even when asked otherwise, missing product context, and plausible but wrong answers (NN/g).


2. Giving AI your voice

A single word like "friendly" makes AI overshoot. Several descriptors work better, and real examples work best (NN/g on tone).

Put this in a context file or at the top of the prompt:

Voice: calm, direct, practical. Never cheerful about problems.
We are: clear (short sentences, plain words), respectful (no blame), useful (always a next step).
We are not: playful, salesy, technical.
Terms: "booking" (never "reservation"), "member" (never "user" or "customer"),
"court" (never "pitch").
Examples of our copy:
- Error: "That slot was just taken. Here are the next free times."
- Empty state: "No bookings yet. Find a court near you."

That one block does more than any clever prompt: it gives the model your voice, your words and your examples (see Few-shot prompting).


3. Useful prompts for content work

Task Prompt to adapt
Variations "Write 8 versions of this button label, max 18 characters, in our voice. Mark the 2 you'd recommend and why."
Error message "Rewrite this technical error for members: what happened, what's still safe, one next step. Max 2 short sentences."
Consistency "Check these 40 strings against our terms list. Flag every inconsistency and suggest the fix."
Plain language "Rewrite this paragraph for a reading age of about 12, keeping every fact."
Tone shift "Make this notification calmer and shorter. Don't add or remove information."
Inclusive language "Review this copy against the checklist below and flag anything that excludes or stereotypes."

4. Reviewing AI copy: a five-point check

Read every AI string as if a colleague you don't fully trust wrote it:

1True?

Facts, numbers, limits and promises match the product.

2Ours?

Voice, tone and terms match the guide.

3Clear?

One idea per sentence; the next step is obvious.

4Kind?

No blame, no jokes about problems, no one excluded.

5Fits?

Character limits, plurals, every language you ship.

A five-point check for every string AI writes. Read it as if a colleague you do not fully trust wrote it. Diagram: Marcus Andrey, CC BY 4.0.
A printed page covered in handwritten edits, insertions and crossings-out, with a pen lying on it.
Editing on paper, the way every AI draft deserves to be read: line by line, by someone who knows what it should say. Photo: Phoebe, cropped, Wikimedia Commons, CC BY-SA 3.0.

5. Translation, localisation and AI

Microsoft's style guide draws a useful line: translation converts the words and is "often automated using machine translation"; localisation adapts the product to a culture and is "done by people who are familiar with the local language and culture" (Microsoft).

AI translation is now very good, but not solved: the 2024 research benchmark of the field was titled "The LLM Era Is Here but MT Is Not Solved Yet", and it still relies on professional human evaluators (WMT24).

Practical rules:

  • Use AI to translate drafts and long tail content; have a native speaker review anything users depend on (payments, legal, safety, onboarding).
  • Give the model a glossary per language, so product terms stay consistent.
  • Avoid idioms in the source: they translate badly (Google style guide).
  • Test the layout with the longest language you support (see the localisation pattern in Prompt patterns for designers).

6. Inclusive language

You can hand an AI an inclusive-language checklist and ask it to review copy against it. Good public sources to adapt:

  • Microsoft's Bias-free communication: gender-neutral terms, people-first language for disability, no "master/slave".
  • Google's Inclusive documentation: also avoid "simple" and "easy", which make people feel bad when something isn't.

AI reflects the text it learned from, including its biases (see Algorithmic bias). It can find problems in your copy; it can also introduce them. Review both ways.


7. Telling users that AI wrote something

Two kinds of text need different rules:

  • Copy you wrote with AI's help (labels, help articles) and reviewed: it's your product's copy. You're responsible for it, as for any copy.
  • Text the product generates live for users (a chatbot answer, an AI summary): users should know. The UK government's service manual requires chatbots to say that answers "do not come from a human and might not be accurate", and that AI information meets the same accuracy standard as other content (GOV.UK Service Manual).

In the EU, the AI Act (Article 50) requires disclosure of AI-generated text published to inform the public on matters of public interest, unless a person reviewed it and someone holds editorial responsibility (Article 50). See AI governance and data protection.


8. Vocabulary

Term Meaning
Voice Your product's personality, stable across all copy
Tone How the voice adapts to a moment (an error vs a success)
Terminology / glossary The approved words for product concepts, per language
Localisation Adapting content and product to a culture, not only translating it
Disclosure Telling people that content was made by AI

9. Try it (no code needed)

  1. Write a short voice block for a product you know (section 2): three descriptors, three "we are not", five terms, two examples.
  2. Ask an approved assistant for 8 variations of one error message, once without the block and once with it.
  3. Review both sets with the five-point check (section 4).
  4. Ask it to translate your favourite into another language you know, and check whether a native speaker would say it that way.

Reflection questions: Which of your product's terms does AI keep getting wrong? Where in your product would AI-generated text need a disclosure?


10. If you lead a team

  • Turn your content guide into a context file every tool can read: voice, terms, examples.
  • Agree what never ships without human review: legal, payments, safety, onboarding, other languages.
  • Decide your disclosure policy for AI-generated text in the product, with legal, before you launch it.

Checklist

Learn more

Books, courses and videos to go deeper

Prompt patterns for designers · Context files · Few-shot prompting · Hallucination · Algorithmic bias · AI governance and data protection · AI for visual design: images and video

Saved only on this device.

↑ ↓ to move · Enter to open · Esc to close