Check an AI-generated image before you publish it

Example result.

Use: decorative open-day header, headline in the left third
Found: extra digit on right person's clipboard hand; lettering-like marks on rack label; leftmost upright stops short of the floor
Discarded: "distorted face" at far left (it's a fire extinguisher)
Decision: fix hand, remove label marks, remove upright; recheck full export
Not approved for: any use that presents it as a photo of our warehouse

This checks quality and fit for your own generated image. It doesn’t tell you how somebody else’s image was made.

1. Put it in the real layout and decide what it claims

Ask what the picture claims. A decorative illustration can introduce a topic. A photorealistic image labelled as your warehouse, team or event says it shows the real thing, and generating a convincing picture doesn’t supply evidence for that. Use a real photograph when readers need to see what exists.

Then place it in its layout with the headline, caption and crop, at normal size and on mobile. Open the original at 100% and zoom into anything doubtful, including the largest version readers can reach.

Why this works: many flaws are invisible in a thumbnail and obvious full size, and some problems are about the claim, not the pixels.

What you should see: the image in place, a clear answer to “what is this image saying?”, and a list of areas to look at closely.

2. Work through seven kinds of detail

Go round the image deliberately rather than taking one general glance. Judge each against the intended style: a simplified cartoon hand is different from a hand merging into a clipboard.

• Hands, limbs and contact. Follow arms to shoulders. Check how hands hold things: extra digits, fingers merging into an edge. A partly hidden finger isn’t automatically missing.
• Text, signs and screens. Read every label character by character. Add essential text separately in the layout, where you can edit it.
• Faces and background people. The main subject usually holds up. The three people in the background often don’t: asymmetric eyes, merged teeth, two people sharing a shoulder.
• Repeating structures. Follow shelves, rails and window frames across the image for supports that stop or patterns that change for no reason.
• Reflections, shadows and lighting. In a realistic scene, check that shadows connect to what casts them.
• Edges and overlaps. Trace where objects meet. Is it clear which is in front?
• Small objects. Glasses, tools, fittings: plausible isn’t the same as correct.

Write each concern as a location and an observation. “Right person’s hand merges with the clipboard edge” is something another reviewer can find. “Looks weird” leaves them the whole image to search.

Why this works: a fixed circuit stops you looking only where the eye is drawn.

What you should see: a short list of findings, each with a location.

3. Ask for a second pass

A colleague can check your list. You can also attach the image in ChatGPT, if you’re allowed to upload it, and ask:

Look at this image for visible quality problems before publication. It's for [where it will appear] at [display size and crop], in a [style] style.
Check hands and limbs, text, background faces, repeating structures, reflections and shadows, edges, and small objects.
For each possible issue, give its location and describe what you can actually see. Separate clear observations from uncertain ones. Don't guess unreadable wording. Don't decide whether the image was made by AI, and don't approve it for publication.

[where it will appear], [display size and crop] and [style] come from step 1, such as “a supplier open-day web page”, “full-width 16:9” and “illustrated”.

Why this works: the categories focus its attention, and asking for locations makes each finding checkable. OpenAI notes that image analysis has limits with counting and precise positions, and images are resized for analysis, so check every finding in the original.

What you should see: a list of locations and observations to compare with yours.

4. Use it, fix it or replace it

• Use it if it suits the purpose, doesn’t mislead and passes your checks.
• Fix it if one thing is wrong and has a clear remedy: a crop, a conventional retouch or a focused edit. Cropping only loses pixels; regenerating risks a new picture with new problems.
• Replace it if it implies an unsupported factual claim, has a consequential error or keeps needing repairs. You don’t have to keep negotiating with the same candidate.

For a focused edit:

Edit this image to correct [problem and location]. The result should be [what it should look like]. Keep the subject, framing and the empty space for text. Don't add lettering, logos or new objects.

Generative edits can change more than you asked for, so go round the seven checks again on the result. Before publishing, check the final export in its layout, including the mobile crop, add any required disclosure, and write alt text (or mark it decorative).

Why this works: the smallest repair keeps the rest of the image as you checked it.

What you should see: a final export that passes the seven checks in the real layout.

Check the result

• Every finding has a location, and you’ve confirmed it in the original.
• Every label and sign has been read, not glanced at.
• The final export has been checked in its layout and mobile crop, after the last edit.
• The image’s claim matches its use: decorative stays decorative.

The usual failure: treating the AI second pass as the check. It can miss a real problem and report one that isn’t there, and “no concerns” doesn’t clear the asset. Spot it by checking every finding yourself, and looking at text and labels even if it didn’t mention them.

Look only at [specific area] in this image. Describe any visible problem and where it is. Separate observations from uncertainty, and don't guess at details you can't see clearly.

[specific area] is the region you want a closer look at, such as “the rack label beside the clipboard”. Supply a crop alongside the full image if you can.

Fictional example. A 16:9 illustrated warehouse scene for a supplier open-day page: three adults stand on the right, leaving the left third clear for a headline. It’s decoration, not a photo of the venue. The reviewer’s notes:

Right person's clipboard hand: an extra visible digit; one finger merges with the board.
Rack label beside the clipboard: marks that look like lettering but don't form words. The brief asked for no lettering.
Leftmost rack upright: stops about fifteen centimetres short of the floor; its shadow continues as though it didn't.
Palette, style and headline contrast: fine in the layout.

The ChatGPT pass found the hand and the floating upright, missed the label entirely, and reported a “distorted face in the background” that turned out to be a fire extinguisher. Three items, one of them wrong. A focused pass on the rack label then described the meaningless marks.

• Cropping wasn’t the answer here. A tight crop round the people would remove the upright but also lose the wide shape and the headline space. Fix the hand, remove the marks and the upright, then check the whole export again.
• The false finding mattered as much as the missed one. Editing out a “face” that was a fire extinguisher would have changed the image for no reason.
• The approval is for this use. If the same image is later proposed as “our warehouse” on a recruitment page, the original approval doesn’t cover it. The claim has changed.

To brief the image in the first place, see write an image brief you can improve one change at a time.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

File and version:
Intended use, sizes and crops:
Findings by location (seven categories, or "not applicable"):
Does the use imply a factual claim? What supports it?
Repairs made, and final export checked:
Decision: use / fix / replace
Alt text or decorative; disclosure:
Reviewer and date:

Sources

• OpenAI Help: ChatGPT image inputs FAQ
• OpenAI Help: Images in ChatGPT
• Google Docs Editors Help: AI labels and Content Credentials
• C2PA: FAQs