The seven AI artifacts shoppers actually notice
Most AI artifacts go unnoticed in casual scrolling. A handful do not — they trigger the "something's wrong" reaction even from shoppers who couldn't articulate why. These are the ones to hunt for during QC.
| Artifact | Where it appears | Severity |
|---|---|---|
| Warped or gibberish text on labels | Bottles, boxes, branded apparel | High |
| Extra/fused fingers or wrong hand anatomy | Held products, model shots | High |
| Asymmetrical product details | Shoes, eyewear, paired items | High |
| Melted or broken seams and stitching | Apparel, bags, upholstery | Medium |
| Impossible reflections or shadows | Glass, metal, glossy surfaces | Medium |
| Repeating-pattern drift | Knits, prints, woven textures | Medium |
| Smoothed or "plasticky" skin | Model photography | Low (but cumulative) |
Text on packaging: the most common AI failure
Text is the single hardest thing for image generators to render correctly. Even leading models still produce subtly wrong characters, broken kerning, or invented logos. On a branded product, that's not a cosmetic issue — it's a credibility problem and a potential trademark one.
Never trust AI-generated text on a product. Always check labels, logos, and packaging character-by-character against the real product, and replace any rendered text with composited real artwork before publishing.
The practical workflow: generate the image with a placeholder or blurred label, then composite the real brand artwork onto the surface. Tools that preserve a product reference photo through generation (rather than fully synthesizing from a text prompt) cut this work dramatically, because the model isn't trying to invent your branding in the first place.
Symmetry failures: the giveaway in pairs
Shoes, earrings, sunglasses, gloves — anything that ships in matched pairs is a high-risk subject. AI models treat the two halves as separate generation problems, and small mismatches creep in: a slightly different heel height, mismatched eyelets, one lens fractionally larger than the other.
The fix is workflow-level: shoot or generate paired items as a single composite where both halves come from the same reference image, rather than two independent generations.
Reflection and shadow logic
Physics is where AI models most often slip in subtle ways that read as "off" without being immediately identifiable. The two most common reflection errors:
- Mismatched reflections. The product's environment in the reflection doesn't match the background. A bottle reflects a window that isn't in the scene.
- Inconsistent shadow direction. The product casts a shadow one direction; nearby props cast theirs another. Even shoppers who can't articulate it sense the wrongness.
For glass, metal, and any glossy material, treat reflection QC as a separate review pass. Look specifically at: where the light source must be based on the shadow, then check that reflections in the product surface match that light position.
A practical QC checklist before you publish
Build this into your image approval workflow. It takes about 30 seconds per image and catches the artifacts that lose sales.
Quick visual scan (5 sec)
- Does anything feel uncanny on first look?
- Is the product instantly recognizable?
- Are colors plausible for the real product?
Targeted artifact check (25 sec)
- Zoom 200% on all visible text — read every character
- Check symmetry on paired or repeating elements
- Verify shadow direction is consistent across the scene
- Look for fused fingers or extra digits if hands are present
- Trace seams and stitching for melted or broken sections
- Check reflections match the implied environment
Review images at 200% zoom on a phone-sized window, not your full desktop. Artifacts that look minor on a 27-inch monitor get magnified when the customer pinches to zoom on the product page.
Choosing tools that produce fewer artifacts in the first place
The cheapest QC fix is generating cleaner images upfront. A few tool-selection criteria reduce artifact rates dramatically:
- Reference-driven generation, not text-only. Tools that condition on a real product photo preserve the actual product shape, branding, and detail rather than reinventing them.
- Product-aware fine-tuning. General-purpose image models hallucinate freely. Tools specifically trained for product photography hold their structure.
- Identity preservation across batches. If you're generating 12 angles of the same item, every output should be the same item — not a sibling.
- Inpainting and edit-in-place. Being able to fix a single artifact without regenerating the whole image saves hours.
Retouchable is built around product-reference generation rather than prompt-only synthesis — the product in your input is the product in your output, which eliminates the largest source of brand-damaging artifacts before QC even starts.