AI for UX · Step 3 · 6 min read
AI design review process
How to keep AI-generated design at your quality bar: six review gates scaled to the risk, a self-review checklist, the failure patterns to train your eye on, and why automated checks are never enough.
AI makes it cheap to produce work that looks finished. That's the danger: a polished screen invites approval before anyone has checked whether it solves the problem, uses the design system, works for every user and handles every state. Review is how you keep the quality bar where it was. AI output is a draft until a designer has reviewed it and signed it off.
What you'll be able to do after reading:
- Run AI-generated work through review gates, scaled to the risk of the change
- Self-review AI output with a checklist before anyone else sees it
- Recognise the seven most common failure patterns of AI design
- Combine automated checks with human judgement, and know why both are needed
1. Why AI output needs more review, not less
Three things change when AI produces the first draft:
- Volume goes up. You can generate ten options in the time one used to take. Reviewing them doesn't get faster.
- Polish comes for free. Spacing, colour and layout look professional even when the idea is wrong. Polish used to be a sign of effort; now it isn't.
- Mistakes look plausible. An invented business rule or a made-up field label reads as naturally as a real one. See Hallucination.
NN/g calls critique the core skill of design in the AI era: when making things is cheap, judging them is the work (NN/g). They also warn that one output is an example, not an evaluation: you only learn what a tool can do by reviewing many outputs, not one good-looking one (NN/g).
2. The review gates
Problem, users, states, rules and components are clear enough to generate from.
Designer + product ownerI understand and agree with every element, and can explain why.
The designer who generated itDoes it solve the problem, consistent with our principles?
PeersReal components and tokens; WCAG 2.2 AA, checked by tools and people.
Designer + design-system ownerDoes it work for real users with real data?
DesignerSpec, design and code agree on behaviour.
Designer + developerScale the gates to the risk:
| Change | Gates |
|---|---|
| A small UI tweak (a label, a spacing fix) | G1 → G3 → G5 |
| A new screen in an existing flow | G0 → G1 → G2 → G3 → G5 |
| A new workflow, or anything touching money, safety or personal data | All six |
3. The self-review checklist (G1)
Run this before anyone else sees AI-generated work.
Intent
System
States and data
Accessibility
Content
Code (if any was generated)
4. Seven failure patterns to train your eye on
| Failure | Looks like | Catch it by |
|---|---|---|
| Generic look | Card grids, big hero numbers, gradients, in a tool that needs density | A principles check: "would our users want this?" |
| System drift | Custom spacing, off-palette colours, one-off components | Design-system and token checks |
| Happy path only | A beautiful default state, nothing else | The states checklist |
| Invented logic | Plausible but wrong rules or fields | Compare with the spec; ask the product owner |
| Fake accessibility | ARIA everywhere, broken keyboard flow | A keyboard test plus an automated checker |
| Demo-ware | Works with 5 rows, breaks with 5,000 | A realistic-data test |
| Same ideas | Every option looks alike | Ask for structurally different options; sketch first |
5. The hidden cost: cleaning up after AI
NN/g describes a "custodial era of UX": teams that ship AI output without review end up spending their time cleaning up inconsistent screens, broken patterns and confused users (NN/g). Review isn't overhead on AI work; it's what keeps AI work from becoming debt.
6. Vocabulary
| Term | Meaning |
|---|---|
| Review gate | A checkpoint work must pass before moving on |
| Self-review | The maker checks their own work first, against a checklist |
| Design critique | A structured review by peers, against the goal and principles |
| Design-system drift | Small deviations that slowly break consistency |
| UX debt | Problems shipped now that cost more to fix later |
7. Try it (no code needed)
- Generate one screen with any approved AI tool from a short brief.
- Run the self-review checklist (section 3). Count the items that fail.
- Look for the seven failure patterns (section 4). Which ones appear?
- Fix the two most serious, then ask a colleague for a five-minute critique without telling them AI made it. What do they notice that you didn't?
Reflection questions: Which gate do you usually skip when you're in a hurry? Which failure pattern would your users notice first?
8. If you lead a team
- Make the self-review checklist part of your definition of done.
- Run a monthly AI output critique: review AI-generated work together and add new failure patterns to the list.
- Spread reviewing across the team so it doesn't all fall to one person.
- Track rework caused by AI output in retrospectives, so the process improves.
Checklist
Learn more
- NN/g: The Core Skill of Design in the AI Era: Critique · One AI Output Is an Example, Not an Evaluation · The Custodial Era of UX: Cleaning Up After AI
- Accessibility: WCAG 2.2 · W3C: selecting evaluation tools · WebAIM Million 2026
Books, courses and videos to go deeper
Related
Design specs for AI · Generating design with Claude Code · Accessibility with and for AI · Design critique · Human-in-the-loop · Definition of Done for design