Skip to content
UX Atlas
Plate AI·009 · AI workflowGrid H6 · Rev. 2026-09-12

AI Design Critique

Use a model as a fast first reviewer against a named framework, then treat every point as a hypothesis rather than a finding.

AI workflowCraft conventionIntermediate

Definition

A model given a design and an explicit evaluation framework can produce a structured critique in seconds. Its value is coverage and speed before a human review, not judgement: it cannot see your users, your constraints or your data.

Craft convention. True because the industry converged on it. Breaking it costs familiarity, not correctness.

On this plate
  1. 1Definition
  2. 2Running it
  3. 3What it is good and bad at
  4. 4In practice
  5. 5Local chart
  6. 6Continue from here

Reach for this when

  • Before a human critique, to clear the obvious.
  • Reviewing your own work when nobody else is available.
  • Stress-testing a design the team agreed on too quickly.

Common mistakes

  • Unframed requests producing generic advice.
  • Measurement claims accepted without checking.
  • A model critique reported as an audit.

Running it

  1. 1

    Give it the context first

    Who the audience is, what the screen is for, the one task that matters, and the constraints you are working under. Without this you get generic advice.

  2. 2

    Name the framework

    Nielsen's heuristics, WCAG 2.2 AA, or your own review criteria. An unframed request produces opinions; a framed one produces a list you can check.

  3. 3

    Ask for severity and reasoning

    Each point should say what breaks, for whom, and how badly. Points that cannot carry that structure are usually filler.

  4. 4

    Ask what it cannot tell

    A direct question about the limits of the assessment surfaces the parts that need real testing, and reduces confident guessing.

  5. 5

    Verify every item

    Check each claim against the actual design before repeating it. Plausible and wrong is the characteristic failure mode.

What it is good and bad at

Do

  • Use it for completeness checks: missing states, unlabelled controls, unhandled errors.
  • Use it to produce a first pass before a critique session, so humans spend their time on judgement.
  • Use it to argue the opposite case, which is useful against a team that has converged too early.

Do not

  • Do not accept contrast or target size claims without measuring. Models estimate these and are frequently wrong.
  • Do not let a model's confident tone substitute for user evidence. It has no access to your users.
  • Do not paste customer data or unreleased material into a service whose retention terms you have not checked.
  • Do not use it as the accessibility audit. It complements a keyboard and screen reader pass and replaces neither.

In practice

  1. i.

    Useful and useless in the same output

    A critique correctly spotted that a form had no visible required-field convention and invented a claim about tap target sizes that measurement contradicted. Both appeared in the same list with the same confidence, which is why the verification step is not optional.

Local chart

AI·009 · Region XII · Grid H6

This entry and the entries it is stated to connect with. Each line is labelled with what the connection means.

Authority BiasBIAS·006 · PRINCIPLES BEHIND THISConfirmation BiasBIAS·009 · PRINCIPLES BEHIND THISScreenshot to UX AuditAI·001 · RELATED CONCEPTUI to Accessibility AuditAI·005 · RELATED CONCEPTHeuristic EvaluationRES·002 · RELATED CONCEPTDesign CritiqueFLOW·001 · RELATED CONCEPT

Continue from here

Each link says what the connection is, so you can tell a principle from an alternative from a thing people mix this up with.

Principles behind this

The reasoning this solution is an application of.

Related concept

Connected closely enough to change how you apply this.