Zoom in and check eight things before any AI image goes out: hands and faces, any text or numbers, logos, light and shadows, objects that merge or defy physics, whether it shows your sector accurately, who is pictured and how, and whether it looks like someone else's work. Then view it at phone size, where most people will see it.
AI image errors are small, and they sit where your eye doesn't go: the edge of the frame, the background sign, the fourth person in a group. At first glance the brain accepts the whole picture and fills in what it expects. Your audience, scrolling past hundreds of images, is surprisingly good at spotting the one with six fingers, and they say so in the comments.
The eight-point image checklist
Each point below says why it goes wrong and how to verify it. Work through them in order; the first four are about the image being physically plausible, the last four about it being right for your business.
1. Hands, faces and bodies
Why it goes wrong: anatomy is still where image models slip most visibly: extra or fused fingers, teeth that run together, mismatched earrings, eyes looking in slightly different directions, an arm with no clear owner in a group shot. How to verify: count fingers on every visible hand, check that each hand belongs to a body, and look at eyes, teeth and ears at 200% zoom.
2. Text, numbers and symbols
Why it goes wrong: models draw letters as shapes. Signs, screens, documents, keyboards, book spines and calculator displays come out as near-words or nonsense. How to verify: read every piece of text in the image, letter by letter, including tiny background text. If you can't read it, neither can your customers, and blurred pseudo-text looks careless.
3. Logos and brand marks
Why it goes wrong: AI either invents logo-like shapes or recreates real brands' marks from its training, sometimes on clothing, cars or products in the background. How to verify: find every logo-like mark. Your own logo should be added afterwards in a design tool, never generated; any other company's mark comes out.
4. Light, shadows, reflections and physics
Why it goes wrong: each part of an image is plausible on its own, but they don't always agree: shadows falling in two directions, a mirror reflecting a different room, a cup merging into a hand, stairs that lead into a wall. How to verify: identify where the light comes from, then check every shadow and reflection against it. Look where objects touch: hands and cups, feet and floors, people and chairs.
5. Your sector's details
Why it goes wrong: the model knows what an office or a house looks like in general, not what your work looks like in particular. A bookkeeping firm's image shows a spreadsheet with nonsense column headings; an insurance broker's "home" image shows a smoke alarm mounted on a wall at knee height. How to verify: have someone who does the work look at it for ten seconds and ask what's wrong. Specialists spot these instantly; everyone else never will.
6. Who is pictured, and how
Why it goes wrong: models default to stereotypes (the adviser is an older man, the assistant a young woman) and generate the same face several times in a group. Faces can also resemble real people, including public figures or your own staff. How to verify: ask who is shown in which role and whether that's the message you want; check group shots for repeated faces; if a face looks familiar, don't use it. The broader checks for bias in AI content are in spotting bias and stereotypes in AI-written content.
7. Resemblance to someone else's work
Why it goes wrong: an image can closely echo a well-known photo or artwork, or carry a ghost of a stock-photo watermark. How to verify: run a reverse image search with Google Lens or TinEye. Near-matches are a reason to regenerate. Check your tool's licence terms allow commercial use on your plan while you're at it.
8. What the image claims
Why it goes wrong: even a flawless image can say something untrue. A polished "our office" picture when you work from a shared space; a "our team" photo of people who don't exist; a product feature you don't offer. How to verify: read the image as a statement next to its caption. If a customer took it literally, would they be misled? If so, change the image or the caption, and see the tutorial on labelling AI-generated images and video for when a label is needed.
Four ways of looking that catch what a glance misses
Knowing what to look for isn't enough; you also need to break the brain's habit of seeing the whole picture. Four techniques, each under a minute:
- Zoom and grid. At 200%, pan across the image in a grid: top-left to top-right, then down a row, and so on. Errors hide in corners and edges because nobody looks there.
- Flip it horizontally. A mirrored image looks new to your eye, and asymmetries you'd accepted (a lopsided face, a crooked door frame) jump out. Painters have long checked their work in a mirror for the same reason.
- Switch to greyscale. Without colour, you see light and shadow directly. Inconsistent lighting that colour disguised becomes obvious.
- View at phone size. Look at it on a phone, in the actual post layout, with the caption. Some errors vanish at that size; others, like garbled text in the middle of the frame, are more noticeable because they're all there is to read.
One image, checked: a filled-in example
An illustrative image for an insurance broker's home insurance renewal post: "a family relaxing in a sunny living room". The checker's record:
| Point | Finding | Result |
|---|---|---|
| 1. Hands and faces | Child's left hand has six fingers; adult faces fine | Fix |
| 2. Text and numbers | Magazine on the table reads "HOME INSRANCE" | Fix |
| 3. Logos | Swoosh-like mark on the father's jumper | Fix |
| 4. Light and physics | Window light from the left, but the lamp casts a shadow to the left too | Minor; acceptable at phone size |
| 5. Sector details | Smoke alarm mounted low on the wall | Fix (an insurer's image shouldn't show bad safety practice) |
| 6. Who is pictured | Fine; no repeated faces | Pass |
| 7. Resemblance | Reverse image search: no near-matches | Pass |
| 8. What it claims | Generic family; caption doesn't claim they're clients | Pass |
Four fixes, all local: the hand, the magazine text, the logo and the smoke alarm could each be corrected with the edit or inpainting tool in the image generator (or removed entirely, in the case of the magazine). About eight minutes of editing, then a quick re-check of the edited areas, because edits introduce their own errors.
An insurance broker's campaign: eight images through the checklist
Scaling the example up. The same illustrative broker generated eight images for a month-long renewal campaign across Facebook, Instagram and LinkedIn. The checks took about six minutes per image, 48 minutes in all, with these results:
- Three passed first time. Simple compositions: a front door, a set of keys on a table, a car on a driveway with no visible number plate.
- Three were fixed with edits. The living-room image above; a kitchen image where the tap merged into the splashback; an office image where a monitor showed a fake insurance form with garbled headings, replaced with a blank screen.
- Two were discarded. One flood-damage image carried the faint diagonal shadow of a stock-photo watermark, which the reverse image search then matched to a real photo agency's picture. The other showed a street that a member of staff recognised as a real local road, with invented damage on real houses, which could upset the people who live there.
Two regenerated replacements took another twenty minutes to create and check. So the campaign's images cost about 70 minutes of checking and fixing. Against that, one pulled post, the replies explaining it, and the question it raises about how carefully the firm handles other things is a far worse outcome. The discarded flood image is the one that matters most: a close match to a real agency photograph is both a rights problem and the kind of thing that ends up on a complaint.
Asking an AI to check the image: useful, not sufficient
Chat assistants that read images (ChatGPT, Claude and Gemini all do) can act as a second pair of eyes. They are good at listing text they can see and obvious anatomical errors, less good at physics and sector details, and occasionally report problems that aren't there. A review prompt to start from:
Check this image for errors before we post it. List:
1. Any text, numbers or letters, exactly as they appear.
2. Any hands, faces or bodies that look malformed.
3. Any logos or brand marks.
4. Anything physically impossible: shadows, reflections,
objects merging.
Describe where each issue is (top left, centre, etc.).
Don't comment on style or composition.An illustrative response for the living-room image before it was fixed:
1. Text on magazine (centre table): "HOME INSRANCE" - misspelt.
2. Child (right): left hand appears to have six fingers.
3. No logos detected.
4. Lighting appears consistent.Compared with the human check, the review caught the text and the hand, which is useful. It missed the logo on the jumper, called the lighting consistent when the lamp's shadow contradicts the window, and could never have known that an insurance broker shouldn't show a smoke alarm at knee height. Use the AI check as the first pass through points 1 to 4, then do points 4 to 8 yourself.
Errors in the details your customers know best
Point 5 deserves examples, because it's where AI images most often embarrass professional firms. Your customers know your world, so they spot these first.
A bookkeeping firm's hero image. Before: a laptop showing a spreadsheet, beside a calculator. At full size, the spreadsheet's column headings read "Totla", "Anount" and "Q5", and the calculator had two "7" keys and no "4". After: the firm regenerated with the screen turned slightly away so no content was visible, and replaced the calculator with a real photo from its own office. The before image would have been seen by exactly the audience most likely to notice figures that don't add up.
A law firm's "our reception" image. A small firm generated a reception area with its name on the wall behind the desk. The sign read the firm's name with one letter doubled, and the reception looked nothing like the real one, which clients visit. A follower pointed out the misspelling within hours. Generated images should never show your own name or your real premises; use real photos for anything customers will compare with reality.
A recruitment agency's "diverse team" image. Eight people around a meeting table, varied at first glance. At 200%, three of them had the same face with different hair. Group shots are where repeated faces hide; count the faces, then compare them.
Prompting to avoid the commonest errors
The cheapest error is the one never generated. Most of the failures above cluster around a few things models draw badly, so compose around them:
- Keep text out of the image. Ask for no text, signs, screens or documents, and add any words afterwards in a design tool, where you control the spelling.
- Keep hands simple or out of frame. Hands holding one clear object, or not visible, fail far less often than hands gesturing or interlocking.
- Limit the number of people. One or two people give the model less to get wrong than a meeting table of eight.
- Generate at the final shape. Ask for the aspect ratio of the platform you're posting to, so cropping later doesn't cut off a head or leave a half-object at the edge.
- Say "no logos or brand marks" and add your own logo afterwards.
An illustrative before and after for the broker's campaign. The first prompt:
A happy family in their living room reading about home
insurance, cosy, sunny, realistic photoIt invited three of the four errors found in the checked image: "reading about home insurance" produced the misspelt magazine, a family meant several sets of hands, and nothing ruled out logos. The revised prompt:
Two adults relaxing on a sofa in a sunny living room, hands
resting in their laps, no text, no screens, no magazines,
no logos or brand marks, realistic photo, 4:5 portraitImages from the revised prompt still went through the checklist, but they passed points 1 to 3 far more often, which is where most of the fixing time had gone.
Fix, regenerate or bin: deciding what to do with a flawed image
Not every flaw needs the same response. A simple rule keeps the time spent proportionate:
| Situation | Action | Typical time |
|---|---|---|
| One or two small local flaws (a hand, a sign, a stray mark) | Fix with the edit or inpainting tool, then re-check that area | 5 to 10 minutes |
| Structural problems (lighting throughout, perspective, several bodies wrong) | Regenerate with a revised prompt | 10 to 20 minutes |
| Close resemblance to existing work, a real person, a real place or someone else's brand | Discard and start again | Whatever a new image takes |
| The image misrepresents your business, even if flawless | Discard, or use a real photo | Varies |
A useful ceiling: if an image has needed more than two rounds of fixes, regenerate or switch to a real photograph. Heavily patched images tend to accumulate small inconsistencies that pass each individual check and still look slightly off as a whole.
Who checks, and when
The person who generated an image is the worst person to check it; they've already accepted it. Build the check into the workflow instead:
- Different person. The creator runs the four viewing techniques; a second person does the eight points. In a very small team, check the next morning with fresh eyes rather than straight after generating.
- Before scheduling, not before posting. Once a post is in the scheduler, people assume it has been checked. Make "image checked" a required field or a tag in whatever tool you use.
- Keep the checklist record. A line per image (pass, fixed, discarded, and why) shows over a month which errors your prompts keep producing. Those go into your image guidelines, covered in brand guidelines for AI images, so the next batch starts better.
Product images need an extra rule: AI can quietly change details of a product it recreates, so show real products in real photos and let AI generate only the setting; the comparison of AI product photos and a photographer covers where each makes sense.
Checking AI images: common questions
Are newer image models accurate enough to skip the check?
No. Newer models make fewer obvious errors, which makes the remaining ones easier to miss: a small sign with one wrong letter, a reflection that doesn't match, a face that looks like someone real. The check gets faster as models improve, because more images pass first time, but the risk of posting an error you didn't look for stays.
How long should checking one image take?
About three to six minutes for a typical social image: a minute zooming through the grid, a minute on text and logos, a minute on people and sector details, and a reverse image search. Images with crowds, text, screens or products take longer. If an image needs more than ten minutes of checking, regenerating it is often quicker.
Do real photos edited with AI need the same check?
A lighter version of it. Generative fill and object removal work on part of the image, so check the edited area and its edges at 200%: textures that repeat, lines that don't join, a shadow left behind by a removed object. Then check that nothing outside the edit changed, especially faces and any text, and keep the original file in case you need to start again.
Further reads
- How to Create Product Photos With AI on a Small Budget — Creating product photos with AI in a way that passes this check.
- How to Create On-Brand Social Media Graphics With AI — On-brand social graphics made with AI.
- AI Alt Text: How to Generate and Check Image Descriptions — Write and check alt text for the images you post.
- How to Check AI Content for Plagiarism Before You Publish — The text equivalent of the lookalike check.
- Can You Legally Use AI-Generated Images in Your Marketing? — Where the legal limits on AI images in marketing sit.
- Social Media Pre-Publish Checklist for AI-Drafted Posts — The check for the words that go with the image.
- AI Mistakes That Damage Customer Trust, and How to Avoid Them — Nine AI mistakes customers notice, why each one stings, how to prevent it, and a 20-minute monthly check that catches problems before customers do.
- AI Model Shots for Boutiques: Try-On Images Without a Photoshoot — How AI on-model images work, what tools like FASHN and Photoroom cost, which garments come out wrong, and how to check and label every image honestly.
- Online Shop AI Mistakes That Hurt Trust and Conversions — Eight AI mistakes that quietly raise returns and lower sales in small online shops, each with a real-looking example and the fix.
- AI Renders vs Outsourced Visualisation: What Practices Save — What a studio still costs per still, what AI render tools cost per seat, and a costed year showing where a small practice really saves.
- Best AI Rendering Tools for Small Architecture Practices (2026) — Eight AI rendering tools for small practices, ranked by how faithful they are to your model, with verified prices, a test brief and when each one wins.
- How Interior Designers Use AI for Mood Boards and Presentations — How interior designers use AI for concept boards, real-product selection boards and client decks, with prompts, sample outputs and the traps to avoid.
- AI for Freelance Designers: Faster Concepts Without Rights Issues — Where AI speeds up concept work, how to pick generators by their terms, a process log that proves authorship, and the checks before you hand over files.
- AI Virtual Staging for Property Listings: Costs and Disclosure — Per-photo and per-listing costs of AI virtual staging, the edits a label can't excuse, label wording to copy, and a pre-publish check for every image.
- Is AI Frame Advice and Virtual Try-On Worth It for Opticians? — What virtual try-on and AI frame advice cost an optician, the frame-digitising cost most people miss, three practices' verdicts and a 30-day test.
- AI Garden Design Visuals in Quotes: Do They Help Close Jobs? — When an AI-edited photo of the client's own garden helps a landscaping quote get signed, when it backfires, and a fair way to test it.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.