Plastic Skin: How an AI-Touched Image Gives Itself Away.
Two years ago, recognizing the failure modes of a generative tool wasn't a professional skill, because the tools didn't produce anything at professional quality. Now they do, and spotting their errors has become part of the job for anyone who edits or commissions images.
The tricky part is that these mistakes almost never look like mistakes. They look like an image that is slightly cheap, without anyone being able to point at why. The client notices, can't articulate it, and comes away feeling the work wasn't worth what it cost.
These are the patterns that keep recurring.
Skin that's too perfect.
The most common and the most expensive. AI tends to smooth toward plastic: it strips texture, strips dimension, and chases an idealized evenness no human skin actually has. The result sits squarely in the uncanny valley and reads as cheap rather than polished.
This is exactly why manual skin retouching didn't lose value to automation. If anything it gained. Understanding how real skin is built is what allows an editor to recognize when the machine has overshot and pull it back.
Shadows that don't add up
Generative fill leaves a trail that becomes obvious once you know what you're looking for. Shadows falling the wrong way relative to the scene's light source. Geometry that doesn't close — an architectural line that bends, a reflection corresponding to nothing. Patterns repeating a little too regularly in textured areas like foliage or fabric.
None of it jumps out at a glance. All of it jumps out at 200%, which is precisely how the art director receiving the file will look at it.
Color that jumps
In batch edits, the tool treats every frame as an isolated case. That produces color and exposure shifts between images that should feel continuous, because they came from the same series, the same minute and the same light.
It's an error you can't catch looking at one photograph. You catch it looking at the full delivery — which is how the client sees it.
Saturation for no reason
There's a structural tendency toward heavily saturated, high-contrast output, because that register performs well in a feed and the model learned from material optimized for exactly that. In real photography, that level of vibrancy rarely appears without a deliberate decision behind it.
When an image looks vivid for no discernible reason, it's worth asking where it came from.
The details that still break
There's a short list of areas where models keep tripping, even as each version improves them slightly.
Hands remain the classic, less because of extra fingers than joints bending where joints don't bend, or two hands on the same person that don't appear to belong to the same body. Any text inside the frame — a sign, a label, a book spine in the background — tends to decompose as soon as you look closely.
Eyes are another weak point: the reflection of the light source should match across both, in position and shape, and very often doesn't.
Transitions are worth checking too. The edge where hair meets background, the point where fabric emerges from shadow, the outline of a pair of glasses. These are areas of high structural complexity, and they're where generation slides toward the average most easily.
The opposite problem
There's a less discussed failure that's starting to cost money: false positives.
Real photographs are being accused of being synthetic. It happens most to highly polished work — controlled studio light, high-end retouching, or unusual atmospheric conditions that genuinely do look impossible. Well-executed beauty portraiture is currently the genre most at risk of someone in the comments insisting the person doesn't exist.
That has two practical consequences. The first is that keeping originals stopped being an archivist's habit and became a defense. The second is aesthetic and rather more uncomfortable: photographers are adjusting their work so it won't read as generated. Dialing back retouching, leaving extra texture, avoiding light that's too clean.
It's a new kind of censorship — soft, hard to detect, and one where the tool shapes the output without having touched it.
How to check without losing your mind
The most common mistake when reviewing is scanning the whole image looking for something odd. That finds almost nothing, because the overall impression is usually correct.
A fixed route works better, always in the same order. Hands and eyes first, since that's where structure breaks. Then any text inside the frame. Then shadow direction against the visible light source. Edges last: hair against background, fabric against shadow, the outline of thin objects.
And on batch deliveries, one extra pass across the thumbnails as a set, which is where color jumps between supposedly continuous frames become visible.
Done in that order it takes under a minute per image and catches most of the problems. Improvising the review takes longer and catches less.
The point
The skill becoming scarce isn't technical execution — automation absorbed most of that. It's judgment: looking at an output and deciding whether it holds.
The editors holding their value are the ones who built themselves a mental catalog of failure modes and check it systematically, rather than trusting that something is good because it looks clean.
Sources: Fstoppers, What a Competent Photo Editor Looks Like in 2026 · Imagine.art, analysis of AI slop markers
Keep reading
Looking for photography made with a camera, with real people and real light? Let's talk about your session.
Comments