ChatGPT for Design Feedback What it can see—and what it cannot verify

The short answer: ChatGPT can help you look, not prove

ChatGPT can analyze an uploaded image and respond to questions about what appears in it. That makes it useful when you need a quick outside read, another interpretation, a way into a stuck decision, or language for a concern you have not been able to name.

It is less reliable when the answer depends on something the image does not contain: the code, interactive behavior, research, organizational constraints, precise color values, or what real users will do. OpenAI's own image-input guidance notes that the model may struggle with precise spatial relationships, graphs, rotated text, counting, and ambiguous images—and may simply describe an image incorrectly.

“AI or human?” is not the most useful question

Human feedback can be vague, preference-driven, or disconnected from the brief. AI feedback can be specific and still be wrong. Instead of choosing a winner, start with the decision and ask what evidence would actually reduce uncertainty.

What ChatGPT can help you examine

Give it a real artifact and a bounded question. Its most useful role is often generative: helping you see more than one reading, prepare for critique, or make the next iteration deliberate.

Research on AI-assisted UI critique supports a measured version of this. In one study of 51 interface mockups, expert designers found AI suggestions useful for some subtle errors, text, and interface semantics, but the usefulness declined across later iterations. Other work found that iterative visual prompting improved the grounding of critique compared with a one-shot response. The lesson is not “AI can critique design.” It is that the artifact, prompt, iteration, and verification process materially change the result.

What it cannot prove from a screenshot

A screenshot is evidence of one visible state. It does not reveal every input, state, rule, or outcome behind the design. Treat the following as questions to investigate—not findings to publish as fact.

Choose the feedback source by the job

Write a prompt that makes uncertainty useful

“What do you think?” invites a plausible performance of expertise. A better prompt gives the model a job, a boundary, and a way to say what it does not know.

1. 1
   Name the artifact and stage

   Say whether this is an early concept, one screen in a flow, a portfolio homepage, a visual identity, or an implemented product.
2. 2
   Share the goal, audience, and constraints

   Include only the context that affects the decision. Do not expect the model to infer a brief it has never seen.
3. 3
   Ask one decision-shaped question

   For example: ‘Does the first screen make the next action and its consequence clear enough for a first-time user?’
4. 4
   Require evidence and options

   Ask it to separate what is visible from what it infers, label uncertainty, and suggest more than one possible response.
5. 5
   Verify before changing the work

   Check the cited evidence, reject invented details, and choose whether the claim needs another person, a test, or a tool.

I am sharing [artifact] for [audience] at [stage].

The decision I need help with is: [one focused question].
Relevant constraints: [only what affects this decision].

First, describe what you can directly observe. Then:
1. Offer two or three possible interpretations.
2. Point to visible evidence for each one.
3. Label anything uncertain or missing from the artifact.
4. Suggest a few options, including a small test when useful.

Do not claim user behavior, accessibility compliance, responsive behavior,
or business impact unless the evidence provided can actually support it.

Where The Crit fits

The Crit is AI-assisted. It is not a professional designer secretly reviewing each submission. Its value should come from a design-specific critique process: capturing the submitted artifact, asking for an evidence-grounded read, checking the response for unsupported claims, and keeping the scope honest.

A free First Read is intentionally focused: it identifies the most useful priority, explains the visible evidence, and gives you a next move. The $9 Full Crit opens the wider review and includes one revision check. Neither replaces research, accessibility testing, or the context of a collaborator who knows the project.

If you are comparing other ways to get feedback, see the practical guide to choosing a design feedback tool by the decision you need to make.

Sources and further reading

• OpenAI: ChatGPT Image Inputs FAQ — supported image tasks and documented limitations.
• Generating Automatic Feedback on UI Mockups with Large Language Models — a study of AI feedback across 51 interface mockups with evaluation by 12 expert designers.
• Visual Prompting with Iterative Refinement for Design Critique Generation — evidence that visual grounding and iteration affect critique quality.
• Google Research: PromptInfuser — a study of tightly coupled UI and prompt iteration with 14 professional designers.
• Google Research: AI is Not Enough — why UI linting benefits from deterministic heuristics alongside language-model flexibility.
• W3C: Web Content Accessibility Guidelines 2.2 — the technical criteria to use when an accessibility claim needs verification.