UX copy seems trivial — "how hard is naming a button?" — until you watch a great writer turn a generic flow into something that feels human, and a sloppy writer turn the same flow into something irritating. AI sits awkwardly in this craft. It's competent for low-stakes copy and dangerous for the parts that matter most.
Where AI is genuinely useful for UX
Field labels and short utility copy. "Email address," "Password," "Confirm," "Save changes." These are functional, conventional, and AI gets them right. There's no creative tension here.
Error message drafts. Given a technical error condition, AI produces clear human-readable copy quickly. Edit for voice but the structure is reusable.
Tooltip and help text. Explanatory short blurbs benefit from AI's clarity discipline. Just give it the technical context plus your tone preference.
Quick variations for testing. "Give me 8 alternative button copy options for a sign-up CTA targeting [audience], in [tone]." Most options are throwaway; one or two surprise.
Empty states for new pages. Generic copy is fine for early-stage drafts of empty states. Refine when the page actually ships.
Where AI fails UX copy badly
Onboarding. First-touch copy where users decide whether your product is for them. AI defaults to safe / generic / friendly-but-bland. The companies with great onboarding (Linear, Notion, Pitch) have a specific voice that AI flattens.
Critical paths and conversion copy. Pricing pages, signup flows, paywalls. The difference between 2% and 4% conversion is the difference between thriving and dying. AI draft is a starting point at best.
Brand voice moments. When users encounter your brand personality (404 pages, success states, milestone moments), AI defaults to default. If your product has any voice — playful, opinionated, premium — AI will neutralize it.
Domain-specific terminology. Healthcare apps, financial products, legal services have specific terms users expect. AI will use approximately-correct words that experienced users immediately recognize as wrong.
Tone shifts. Apologetic in errors, celebratory in success, serious in checkout. AI tends toward middle-of-the-road tone everywhere, which sounds robotic in moments where the human would naturally adjust.
A workflow that actually helps
Step 1: write a voice doc. Document your product's voice with concrete examples. "We sound like a competent friend, not a cheerleader. We say 'looks like' instead of 'we encountered an error.' We never use exclamation points. We avoid 'awesome,' 'amazing,' 'totally.'"
Step 2: feed the voice doc to AI for every UX copy task. Don't ask AI to write copy without context; ask it to write copy in your voice. The voice doc is the input that makes outputs usable.
Step 3: AI for variations, you for selection. AI is great at generating 10 options. You're better at picking the one that fits.
Step 4: never ship AI-direct on hero moments. Hand-craft hero copy: landing pages, onboarding, paywalls, brand-defining surfaces.
Step 5: test. A/B test conversion-critical copy. AI guesses well but doesn't know your users. Real users tell you which copy works.
Common AI UX copy mistakes
- Excessive friendliness. Smileys, exclamation points, "Hooray!" Most products don't sound like this. Strip them.
- "Don't worry, [reassuring thing]." AI loves this construction. It's condescending. Replace with the actual information.
- Generic placeholders. "Enter your name here" instead of useful examples like "Tao Xu."
- Apologizing for the user's actions. "Oops! Looks like that didn't work." The system messed up; don't blame the user.
- "In order to" instead of "to." AI uses formal constructions when terse is better.
- Inconsistent capitalization. AI shifts between Title Case and sentence case across an interface. Pick one and enforce.
Brand voice guards
If you have an established voice, build prompts that protect it. The pattern:
Write UX copy for [scenario]. Voice rules:
- We never use [list banned words]
- We always [specific stylistic choice]
- We sound like [reference: a brand or person whose voice you match]
- Never use exclamation points
- Sentence case for all UI labels
Produce 5 options, ranked by best fit.
The more specific the voice rules, the better the output. Vague guidance ("be friendly") fails. Specific guidance ("sound like Linear's product copy") works much better.
When NOT to use AI for UX copy
For brand-defining hero copy. Pay a writer or write it yourself. The 5% of copy that hero appears occupies 50% of brand impression.
For copy that needs domain expertise (medical, legal, financial). AI doesn't know your industry's regulatory constraints; mistakes here are liability, not just style.
For multilingual voice maintenance. AI translates UX copy but flattens voice across languages. If your product has personality in English, you need a multilingual voice strategy, not just AI translation.
When to question your UX copy approach entirely
If your product needs constant copy changes to convert — UI churn, repeated A/B tests on every microcopy — the underlying flow may be wrong. UX copy can clarify a working flow; it can't fix a broken one.
Decision tree
- Functional UI labels, error messages, tooltips: AI with voice doc
- A/B test variations: AI for options, human for selection
- Onboarding, pricing, paywalls: Human-written, AI for sanity-check variations
- Brand-defining hero copy: Pay a writer
- Domain-specific (medical, legal): Human + domain expert
Next steps
- Write your product's voice doc (1-2 pages, examples-first)
- Build a prompt template that injects the voice doc + task
- Run an A/B test on AI vs human-written copy in a low-stakes flow
- Read about UX writing as a discipline (Microcopy, Strategic Writing for UX)