Why AI headshots look slightly off, and which tells are fixable

There is a particular look that makes people pause on a profile photo. Not “that is fake”, just a small hesitation. The face is fine. The lighting is nice. Something is wrong and nobody can name it.

The tells are nameable, though, and they fall into two groups. Some come from a model drawing a face that never existed. Others come from ordinary retouching turned up too far, which means they happen to real photographs too, and those are the ones you can actually fix.

The tells that come from generation

Skin with no pores. Real skin has texture at every scale: pores, fine lines, slight unevenness of colour. Generated skin often has the large-scale shading right and nothing underneath it. The face looks correctly lit and made of something that is not skin.

Symmetry that is too good. Nobody’s face is symmetrical. One eye sits slightly higher, one side of the mouth lifts more, the nose is a little off centre. Models trained on averages drift toward the average, and a face that is too balanced reads as a face nobody has.

Ears that do not match. Ears are complicated, usually partly hidden, and the part a model has least information about. Mismatched ears, or one ear with anatomy that does not quite work, is among the most reliable tells there is.

Hair that stops rather than ends. Real hair has flyaway strands and a soft transition into the background. Generated hair often has a suspiciously clean outline, or strands that merge into one another in a way that does not resolve when you look closely.

Clothing that does not continue. Collars that change direction halfway, a lapel that becomes a different lapel, a necklace whose chain does not meet the pendant. The face gets the model’s attention and the clothes get less of it.

Eyes not looking at the same thing. Subtle, but if the two irises are not consistent with a single gaze direction, people feel it immediately, even at small sizes.

Backgrounds that do not survive attention. Blurred is easy. A generated background often contains objects that are almost something: a shape that is nearly a window frame, a plant that is not quite a plant.

The tells that come from over-retouching

Worth separating, because these appear on genuine photographs that have simply been pushed too hard, including ones processed by tools like this one if you run the sliders to the end.

Plastic skin. Aggressive smoothing removes texture along with blemishes, producing the same poreless look as generation, from a different cause. It is why the smoothing here is edge-aware and hands a portion of the fine detail back afterwards. Without that step, strong smoothing looks synthetic quickly.

Teeth that glow. Whitening applied as brightening rather than as removing yellow. Teeth are not white, they are a pale warm grey with shading between them. Teeth brighter than the whites of the eyes always look wrong.

Over-sharpened eyes. A halo around the iris, or eyelashes that look drawn on.

Under-eyes erased completely. Everyone has some shadow there. Removing all of it flattens the face into a mask and tends to make people look older rather than younger, because it removes structure the eye expects to find.

Hair recoloured as a block. Dyeing every hair pixel the same tone removes the light and dark variation that makes hair read as hair. Even single-colour hair varies enormously strand to strand.

Halos around the head. A bright or dark outline where a subject was cut from one background and dropped on another. It means the matte was too generous or too tight and nobody checked the edges.

The useful part: the second list is fixable

If a photo of yours shows anything from the second group, turn the effect down rather than abandoning the photo. Most tools, including this one, default to moderate settings for exactly this reason. The failure mode is nearly always somebody pushing a slider to the end because more felt better at the time.

A good habit: process it, walk away for ten minutes, then look again. The over-processed version always looks fine while you are doing it and obvious an hour later.

Compare against the original at the same size, too. If your eye keeps going to the skin rather than to the face, the skin has been done too much.

Why this site is built the way it is

That whole distinction is why the pipeline here does not generate anything.

Two small models work out where things are: which pixels are background, hair, skin and clothes, and where 478 points on the face sit. Everything after that is ordinary image processing inside those regions. Your face is never redrawn, so it cannot drift toward an average, the ears stay the ones you have, and the asymmetries that make you recognisable survive.

The cost is real and worth stating plainly: it cannot put you in different clothes, it cannot fix focus, and it cannot turn a profile view toward the camera. We went through that trade in are AI headshots OK for LinkedIn.

To see what your own photo looks like handled this way, try it. Then run the comparison above: put the result beside the original and check that the freckles, the hairline and the slightly uneven smile are all still where you left them.

Frequently asked questions

Can people tell if a headshot is AI generated?

Often, though usually as a vague sense rather than a confident judgement. The most reliable tells are texture-free skin, mismatched ears, too much symmetry, and hair with an unnaturally clean outline.

Why does AI skin look fake?

The fine texture is missing. Real skin carries detail at a scale below the shading, and both generated images and heavily smoothed photographs tend to lose it.

How do I stop my retouched photo looking artificial?

Moderate settings, keep skin texture, never brighten teeth past the whites of the eyes, and leave some shadow under the eyes. Then look again an hour later before uploading.

Do generated headshots keep getting better?

The obvious tells keep receding. The identity question does not change with quality, though: a better-drawn face is still a drawn face rather than a photograph of you.

What is the safest option for a work profile?

A recent photograph of you with the background, lighting and crop improved. That captures almost all of the visual benefit with none of the recognisability risk.