10 Signs an AI Headshot Looks AI-Generated (and How to Avoid Each)

An AI headshot usually looks fake for one of ten reasons: waxy skin, lifeless eyes, warped glasses, mangled jewelry, impossible lighting, melted backgrounds, bad hands, unnatural symmetry, drifting likeness, or uneven detail. Most are visible at 100% zoom in under a minute — and most are avoidable at the input stage.

Why AI headshots fail in the same predictable places

A diffusion model does not photograph a person. It starts from noise and repeatedly nudges pixels toward something that matches a prompt and a learned identity. That process is excellent at the statistically common parts of a portrait and unreliable at anything rare, small, reflective, or governed by physics.

There is no 3D scene inside the model, no light source, no lens. It has learned what photographs of lit faces tend to look like, and it approximates that appearance. So the center of the face — where nearly every training image puts detail — comes out convincing, while the ear, the earring, the eyeglass hinge and the window behind you get whatever attention is left over. If you want the mechanics in more depth, we covered them in how AI headshot generators actually work.

Almost every artifact in this article falls into one of three buckets. Knowing which bucket you're looking at tells you whether the photo is fixable or should be thrown away:

• Physics failures — light, shadow, reflection and refraction that could not coexist in one room. Usually unfixable without a full regeneration.
• Rarity failures — small, uncommon, high-detail objects: jewelry, hinges, teeth, buttons, text. Sometimes croppable, sometimes retouchable.
• Identity failures — the image is a competent photograph of someone who is almost you. Never fixable in post; it's a data problem at the input stage.

1. Waxy, pore-less skin

The most common tell is skin that looks poured rather than grown. Real skin has pores, fine lines, uneven color, a little shine on the forehead and nose, and different texture on the cheek than on the chin. Over-smoothed AI skin has none of that. It reads as candle wax.

Two things cause it. Training sets lean heavily on retouched commercial and beauty photography, so the model's idea of "professional headshot" is already airbrushed. Then the final upscaling pass often re-renders fine texture as a smooth gradient, wiping out whatever pore structure survived. The problem compounds if your input selfies were already smoothed by your phone's beauty processing — you're teaching the model that your face has no texture.

• Turn off beauty mode, skin smoothing and "portrait enhance" before shooting your input photos.
• At 100% zoom, look at the cheek just below the eye. If you cannot see pore structure, reject the image.
• Check for specular highlights — a real face has slight shine on the forehead, nose bridge and chin. Uniform matte across the whole face is synthetic.
• If you retouch, add fine grain rather than sharpening. Sharpening a waxy face produces plastic edges, not texture.

2. Dead eyes and mismatched catchlights

Eyes are where viewers decide whether a face is alive. In a real photograph both eyes carry the same catchlight — the reflection of the light source — in the same position, the same shape and roughly the same size. Generated portraits frequently get one eye right and improvise the other.

Look for catchlights at different clock positions in the left and right eye, or one eye with a catchlight and one without. Then check the pupils: they should be the same size and centered the same way. Common secondary tells include irises whose radial fibers form a repeating decorative pattern, a sclera that is suspiciously clean and uniformly white with no visible vessels, a lost lower lash line, and gaze axes that diverge by a degree or two so the person appears to look slightly past you.

• Zoom to 200–400% on each eye separately and compare the catchlights side by side.
• Reject images where the two eyes disagree about where the light is.
• Favor outputs with a visible lower lash line and a faint shadow in the eye socket — flat, shadowless eyes read as dead.
• Include input selfies taken near a window, so the model has genuine catchlight data to learn from.

3. Glasses that break the rules of optics

Glasses are a physics problem the model does not actually solve. A real lens refracts: the edge of your face behind the lens shifts slightly, the frame casts a small shadow on the cheek, the arms run straight back to the ears, and there is usually a reflection somewhere. Generated glasses tend to be decorative overlays.

The specific failures are consistent. Frame thickness changes across the bridge. One arm disappears into the hair at an impossible angle, or reappears on the wrong side of the ear. There is no refraction offset where the frame edge crosses your cheekbone — the face behind the lens lines up perfectly with the face outside it, which real optics would not allow unless the lenses are plain glass. Lens reflections point at a light that does not exist elsewhere in the frame. Sometimes one lens is rimless and the other is not.

• Trace each arm from the lens to the ear. If you lose it, the image is broken.
• Look at the line where the frame crosses your face and check for a slight offset — its total absence is a tell, but so is a wild one.
• Supply plenty of input selfies wearing the exact glasses you want in the output, from several angles.
• If your prescription frames keep failing across generations, generate without them rather than accepting broken geometry.

4. Jewelry, collars and small hardware that don't survive

After eyes and glasses, the next thing to inspect is every small manufactured object in the frame. Earrings, chains, buttons, collar points, lapels, zips, lanyards and watch faces are low-frequency in training data and high in detail, which is exactly the combination the model handles worst.

Typical damage: one earring rendered cleanly and the other as a soft blob; a necklace that vanishes behind a collar and reappears at the wrong height or the wrong thickness; a chain that clips through skin instead of resting on it; collar points of different shapes and lengths; a blazer lapel that is wider on one side; stitching that runs somewhere and then stops; buttons that don't align with buttonholes. Any embroidered logo or badge text will usually be dream-language — legible-looking letterforms that spell nothing.

• Inspect the left and right side of the collar as separate objects, not as one garment.
• Follow every chain or necklace along its whole visible path.
• Treat any text in the image — badge, lanyard, embroidered logo — as an automatic reject unless it is perfectly correct.
• Choose simple, structured clothing and minimal jewelry for the input set. Fewer small objects means fewer chances to fail.

5. Lighting that could not exist

In a real photograph, three things agree with each other: the position of the catchlight in the eyes, the direction of the shadow under the nose and chin, and the way the background falls off in brightness. When a generated image is wrong, it is usually because those three were rendered independently and never reconciled.

Run the sequence deliberately. If the catchlight sits at ten o'clock, the light is up and to the subject's right, so the nose shadow should fall down and to the left and the left side of the jaw should be slightly darker. Now check the background: if it is brightest on the opposite side, the scene has two key lights facing each other, which almost never happens in a headshot. Watch also for a rim light along the shoulder or hair with no visible source, a chin with no shadow underneath it at all (the "floating head" look), and a subject whose overall brightness does not match the room they are standing in.

• Identify the light source from the eyes first, then verify the nose and chin shadows agree.
• Compare the subject's exposure to the background's. A person lit like a studio dropped into a dim office is the classic composite giveaway.
• Look for a shadow under the chin and along the neck. Its complete absence is a strong signal.
• Prefer outputs with one dominant light and soft falloff over ones lit evenly from everywhere.

6. Background melt

Backgrounds get the least of the model's attention, so they are where coherence dissolves first. The blur can look plausible at thumbnail size and turn into a semantic soup at full resolution: shapes that are almost office furniture, almost windows, almost bookshelves, but that stop making sense the moment you follow a single line across the frame.

Specific things to hunt for: window frames whose horizontal bars do not line up on either side of your head; bookshelf spines with confident but meaningless text; plant leaves that fuse into one another; ceiling or wall junctions that bend as they pass behind you; a repeating micro-pattern in the bokeh; and depth of field that does not increase with distance, so an object three meters away is as sharp as one at one meter. Also inspect the boundary between hair and background — a faint bright halo, or hair strands that end in a clean arc, means the subject and the scene were never in the same space.

• Pick one straight line in the background and follow it across the head. If it jumps, discard the image.
• Zoom on the hair edge and look for halos or unnaturally clean cutouts.
• Choose plain or gently graded backgrounds over busy office scenes — fewer objects, fewer contradictions. Our headshot background guide covers which one suits which purpose.
• Be suspicious of any legible text behind you.

7. Hands, shoulders and the edge of the frame

The six-fingered horror shows of 2023 are largely gone, but hands still fail quietly — and they fail most often in exactly the poses people request for headshots: arms crossed, hand near the chin, one hand in a pocket. Shoulders and the crop edge deserve the same scrutiny, because they are usually the last thing anyone looks at.

Check finger count, obviously, but also check joint direction, knuckle spacing and whether the thumb is on the correct side of the hand. Then move outward: shoulder slopes that are wildly asymmetric, a neck that meets the shoulders slightly off-center, a jacket that changes texture or color between the shoulder and the bottom edge of the crop, and an arm that exits the frame at one angle and could not plausibly connect to anything. The safest fix is structural rather than cosmetic.

• Prefer head-and-shoulders crops. If no hands are in the frame, hands cannot be wrong.
• If you want a half-body shot, count fingers and check thumb placement at full zoom before anything else.
• Scan the bottom 10% of the image separately — it gets the least attention from both the model and the viewer.
• Crop tighter rather than retouching a broken hand. It is faster and more convincing.

8. Over-perfect symmetry

Real faces are asymmetric. One eye usually sits a hair higher, one eyebrow is shaped differently, the smile pulls harder to one side, the nose leans slightly. When a model averages across a training set it can smooth those quirks away, producing a face that is mirror-perfect and, for reasons most viewers can't name, unsettling.

The same over-regularization shows up in details around the face. Teeth come out identically sized, evenly spaced and uniformly white, with no visible gum line variation. Hair has no stray strands, no flyaways, no gaps at the part. Eyebrow hairs run in tidy parallel rows. Ears match each other exactly, which real ears almost never do. Individually each of these is minor; stacked together they produce the plastic, mannequin quality people describe as uncanny.

• Flip the image horizontally. If it looks essentially identical, it is too symmetric.
• Check the teeth: real ones vary in size, shade and alignment.
• Look for at least a few stray hairs. Perfect hair is a rendering artifact, not grooming.
• Include unposed, candid input selfies so the model learns your actual asymmetries rather than a generic average.

9. Likeness drift

This is the most dangerous artifact because the photo is technically flawless. The skin has pores, the lighting is coherent, the background holds up at 400% — and it still isn't you. Distinctive facial geometry has quietly regressed toward the population average, and the person in the frame is a close relative rather than you.

The pattern is consistent enough that identity-personalization research treats it as the central problem: methods like HyperLoRA are explicitly built to hold facial identity stable while the rest of the image varies, because faster zero-shot approaches tend to trade fidelity for flexibility. In practice, skin character, moles, hair and lighting transfer well, while strongly distinctive geometry — an unusual nose, wide or narrow eye spacing, an unusually long or short face — drifts toward typical proportions. Layered on top of that is a documented beauty and age bias in generative image models: analyses of AI-generated portraits have found systematic smoothing toward conventional attractiveness and youth, with sharper jaws, fewer wrinkles and straighter teeth than the input warrants.

There is a practical bar for how much drift is acceptable, and it is not aesthetic. A LinkedIn spokesperson told CNBC that the platform does allow tools, including AI, to enhance or create profile photos, but that "the photo must reflect your likeness" — and that profile photos which don't comply with its policies may be removed. The working test is whether someone who meets you at a conference or joins your video call would recognize you from the photo. If not, you have a compliance problem as well as a credibility one — a topic we go through in is it OK to use an AI headshot on LinkedIn.

• Don't judge likeness yourself. You see your own face mirrored and will miss drift that others catch instantly.
• Send three or four candidates to people who know you well, with no context, and ask which looks most like you.
• Discard any image where the jawline, nose or apparent age is noticeably different from a recent unedited photo.
• Feed a wider range of input selfies — different days, angles, expressions and lighting — so the model has less room to guess.

10. Uneven detail density and the "AI sheen"

A real lens has one plane of focus, and sharpness falls off predictably in front of and behind it. Generative models distribute detail by perceived importance instead of by distance, which produces images where the sharpness map makes no optical sense — a razor-sharp collar next to a soft ear that sits on the same plane.

The related tell is a global glow that photographers sometimes call the AI sheen: mildly HDR-looking, shadows lifted, contrast compressed, everything slightly luminous and nothing quite dark. Real portraits have genuine black points and genuine specular highlights. If you open the file and lift the shadows aggressively, a real photograph reveals sensor noise with a consistent grain structure; generated images often show either nothing at all or a scattering of odd, patterned artifacts. This is a soft signal rather than proof, but it is quick to run and it catches heavily processed output.

• Pick two objects at the same distance from the camera and compare their sharpness. Big mismatches mean it isn't a photograph.
• Check that hair, fabric weave and skin texture all soften at a similar rate as they recede.
• Look for a true black somewhere in the frame and a true highlight somewhere else.
• Raise the shadows in any editor and look at what's hiding there.

The 90-second quality checklist

Run this on every generated photo before it goes anywhere public. It takes about a minute and a half at full zoom and catches the overwhelming majority of artifacts. Work in this order — the early steps kill bad images fastest, so you rarely have to finish the list.

1. 1Eyes. Compare catchlights in each eye: same position, same shape. Check pupil size and gaze alignment.
2. 2Skin. At 100% on the cheek, confirm visible pore texture and some specular shine.
3. 3Light logic. Derive the light direction from the eyes, then verify the nose shadow, chin shadow and background brightness all agree.
4. 4Glasses. Trace both arms to the ears. Check frame thickness and lens symmetry.
5. 5Small hardware. Compare earrings, collar points and lapels left versus right. Follow any chain along its full path.
6. 6Text. Any letterform anywhere in the frame must be perfectly correct or the image is out.
7. 7Background. Follow one straight line across the head. Check the hair edge for halos.
8. 8Hands and crop edge. Count fingers, check thumbs, scan the bottom of the frame.
9. 9Symmetry. Flip the image horizontally; it should look meaningfully different.
10. 10Likeness. Show it to two people who know you and ask, without prompting, whether it looks like you.
11. 11Sanity check at thumbnail size. View it at 100 pixels wide, which is how most people will actually see it. Some images that survive full-zoom inspection fall apart when shrunk, and vice versa.

What the file itself says: Content Credentials and SynthID

Visual inspection is not the only signal anymore. Two provenance systems now sit underneath a growing share of images, and both are worth understanding — partly so you know what a recruiter or a journalist could check, and partly so you know how little any of it currently proves in either direction.

C2PA Content Credentials attach signed metadata describing who made an image, with what device or software, and whether AI was involved. Hardware support is real and growing: the Leica M11-P shipped it first, Canon's EOS R1 and EOS R5 Mark II support it, Sony has enabled it across several Alpha bodies by firmware, and Google's Pixel 10 signs photos with Content Credentials at the point of capture by default. In May 2026 Canon announced a C2PA-based Authenticity Imaging System aimed at news organizations, rolling out first across Europe, the Middle East and Africa for those two camera bodies.

Google's SynthID takes the opposite approach: an imperceptible watermark embedded in the pixels themselves at generation time, which survives screenshotting and re-uploading in a way metadata does not. Google opened a SynthID Detector portal to journalists and researchers at I/O in May 2025, and by I/O in May 2026 reported that SynthID had tagged more than 100 billion images and videos, alongside a rollout of SynthID and C2PA verification inside Google Search and Chrome. OpenAI moved the same way on 19 May 2026, becoming C2PA-conformant and adding SynthID watermarks to images from ChatGPT and its API, with a public checking tool at openai.com/verify.

Two honest caveats. First, most social platforms strip or rewrite metadata on upload, so the absence of Content Credentials proves nothing at all about an image. Second, detector coverage only extends to participating models, and consumer-facing "AI detector" websites are unreliable in both directions — they flag real photos and clear synthetic ones often enough that you should not treat their verdict as evidence. Your eyes at 400% zoom remain the more dependable tool.

When to fix, when to regenerate, and when to book a photographer

Not every artifact deserves the same response. Some are five minutes of retouching, some are a regeneration with different inputs, and some mean an AI headshot is the wrong tool for this particular face, wardrobe or use case. Being honest about which is which saves a lot of wasted effort.

Worth fixing

• Slightly waxy skin — adding fine grain and dialing back local contrast is usually enough.
• A single soft background object — clone or crop it out.
• Minor crop-edge weirdness — tighten the crop.
• Overall AI sheen — restore a black point and reduce the lifted shadows.

Regenerate instead

• Any eye problem. Repainting catchlights convincingly is harder than it looks and usually makes things worse.
• Broken glasses geometry.
• Contradictory lighting, which is a whole-image failure, not a local one.
• Mangled jewelry that sits centrally in the frame.

Consider a real photographer

If likeness drift persists across multiple generations, if you need the exact frames, uniform, hijab, prosthetic or medical device you actually wear rendered faithfully, if you need full-body or environmental shots, or if you're shooting for a context where provenance genuinely matters — press, legal, medical, expert-witness work — a human photographer is still the more reliable answer. They also give you something no generator does: real-time direction, and a negotiation about how you want to be seen. We compare the two paths in detail in AI headshots vs. a photographer, and benchmark what studios actually charge in our 2026 headshot price guide.

How to prevent most artifacts before you generate

Most of the problems on this list originate in the input set, not the model. Generators reproduce what they are shown, including your phone's skin smoothing, your single fixed head angle, and the one pair of glasses you happened to be wearing in every photo. Better inputs eliminate whole categories of artifact for free.

• Vary the angle and the day. Ten selfies from the same afternoon at the same angle teach the model one narrow view of your face and invite drift everywhere else.
• Turn off every enhancement. Beauty mode, skin smoothing, filters and heavy portrait-mode blur all strip the texture the model needs.
• Shoot in soft, directional light. A window at 45 degrees gives clean catchlights and readable shadow structure. Overhead office fluorescents give neither.
• Include the accessories you actually want. If you want your glasses in the output, wear them in most of the inputs, from several angles.
• Keep the wardrobe simple and consistent. Structured collars, minimal jewelry, no busy patterns or logos.
• Don't crop your face tight. Leave head and shoulder room so the model sees how your head sits on your neck.

We go through the full input spec — resolution, count, angles, expressions and the specific photos to avoid — in photos for AI headshots. It is the highest-leverage half hour in the whole process, because no amount of inspection at the output stage can recover detail that was never in the input.

Does it actually matter if people can tell?

It matters, but not in the way most people assume. The evidence suggests the problem is rarely aesthetic — it is about disclosure and about the gap between the photo and the person. Undetected AI headshots do fine; detected-and-undisclosed ones do badly.

Ringover surveyed 1,087 US recruiters in 2024 and found the pattern clearly. In a blind test with no context, 76.5% preferred the AI-generated headshot over the real photograph, and recruiters correctly identified which was AI only 39.5% of the time, despite 80% believing they had been accurate at spotting them. Generator quality mattered: headshots from free tools were caught most of the time, while output from mid-range and top-tier tools was taken for real photography around 60% of the time. But two-thirds of the same recruiters said they were put off once they learned a headshot was AI-generated, and 88% thought candidates should disclose it.

The academic picture points the same way. A Lancaster University-led study published in the Journal of Vision in July 2026 asked 169 participants to sort 96 faces into real and synthetic, and they managed 58.4% accuracy — barely above chance. Worse for anyone hoping realism is the whole story, diffusion-generated faces were rated more trustworthy than real ones: 4.70 versus 4.03 on a seven-point scale. The discomfort people report about AI headshots clearly isn't coming from how the images look.

The practical read: technical polish gets you past the eye, but it doesn't get you past the moment someone meets you and the photo doesn't match. That is why likeness drift belongs at the top of your rejection criteria and waxy skin near the bottom. A slightly imperfect photo that unmistakably looks like you beats a flawless one that doesn't — every time, in every context that ends with a human conversation.

Run the checklist. Reject aggressively. Then look at the winner and ask the only question that really counts: would someone who has met you once pick you out of a room from this photo? If yes, use it. If no, generate again.