People rarely say “that headshot is AI.” They say “something’s off.” Then they scroll past. The tell is almost never the thing you would expect - it isn’t resolution, and it isn’t the background. It’s one of four specific failures, and each one has a cause you can see in the source photos before you ever hit generate.
1. Skin that has never had a bad night
The most common giveaway is retouching that removes the evidence a face is alive: pores, the faint asymmetry under the eyes, the line that shows up when someone actually smiles. Beauty filters have trained everyone to spot this in under a second, because for a decade it has meant this image has been worked on.
A real headshot from a real photographer keeps texture. Skin has a slight sheen on the forehead and nose, hair has stray strands that don’t follow the rest, and there is a shadow under the chin. When all of that is gone, the brain reads mannequin, and it reads it faster than it reads the face.
What to check: zoom to 100% on the cheek and forehead. If you can’t see skin texture - not blemishes, just texture - the image is over-processed regardless of how flattering it looks.
2. A face that is nearly yours
This is the failure that matters most and gets discussed least. Many tools don’t render your face - they render a plausible face that resembles your reference photos. The jaw gets a little squarer, the nose narrows a couple of percent, the eyes move a millimetre apart. Every change is defensible on its own; together they produce a person who could be your sibling.
It doesn’t matter that the result is attractive. A headshot has one job: the person who meets you afterwards should recognise you. A colleague seeing your profile should not have the flicker of wait, is that him? And on a dating profile, a face that isn’t quite yours isn’t just uncanny - it sets up a first date that starts with disappointment.
This is a deliberate design choice on our side. Phoxel’s portrait generations are configured to preserve the likeness in your reference photos rather than invent a more marketable version of it, which is also why we don’t offer face-swapping onto other bodies or scenes. A headshot generator that improves your bone structure has solved the wrong problem.
What to check: open the result next to a recent photo of yourself, at the same size, side by side. Look at the width of the jaw, the shape of the hairline, and the distance between the eyes. Not “does it look good” - “is it the same person.”
3. Everyone gets the same office
Generated backdrops collapse toward the average of their training data: a grey-blue gradient, a suggestion of window light, an out-of-focus plant. Once you’ve seen three of them you’ve seen all of them, and anyone who has scrolled a professional network has seen three hundred.
The same happens with clothing. Ask for “business professional” and you get the same navy blazer over the same white shirt, on everyone, which is exactly how a set of headshots announces itself as generated when the whole team posts theirs the same week.
What to check: for a team set, look at the photos as a grid, not one at a time. If the backgrounds are interchangeable and the outfits rhyme, vary the setting per person - that’s the point of picking a style rather than accepting the default.
4. Light that doesn’t come from anywhere
Real light has a source. It falls off in one direction, it puts a catchlight in the eyes at a consistent position, and it casts a shadow the same way on the face, the neck, and the collar. Generated portraits often blend two or three lighting setups: soft light on the left cheek, a rim light on the right that no lamp in the room could produce, and eyes with catchlights pointing in different directions.
Most people can’t name this when they see it. They just feel that the photo is somehow flat, or somehow theatrical.
What to check: find the catchlight - the bright dot in the iris. It should be in the same position in both eyes. Then confirm the shadow under the chin falls away from that same source.
What actually decides the outcome
Three of these four failures are set before generation, by the photos you upload. A generator cannot know what your face looks like in even light if every reference is a phone selfie shot from below in a dark room; it will fill the gaps, and filling gaps is exactly where likeness drifts.
Two or three straight-on photos in daylight, taken at different moments rather than three frames of one burst, do more for the result than any style setting. We wrote the details in three reference photos that decide everything - the same rules that apply to a product shot apply to your face, with one addition: don’t send only photos where you’re already smiling in the same way. It teaches the model one expression.
The five-second check before you post it
Put the result and a recent real photo side by side and answer four questions:
- Can I see skin texture at 100%?
- Is the jaw, hairline and eye spacing the same as in the real photo?
- Would this background look identical if a colleague generated theirs?
- Are both catchlights in the same place?
If the answer to all four is yes, no one is going to squint at your profile. If the answer to the second one is no, nothing else matters - regenerate it, because that’s the one people notice without knowing why.