How to Choose the Best AI Headshot From 50 Results

Choose the best AI headshot by running every image through five tests in order: likeness, expression, framing, background fit, and technical quality. Shortlist twelve to fifteen candidates fast, then cut to three finalists. The winner is the shot a stranger would still recognize you from — and would happily take a meeting with.

Why the pick matters more than the batch

The batch isn't the product; the pick is. Research on everyday photographs found that different images of the same person vary so widely that strangers frequently sort them into two or more separate identities. Which single frame you publish changes how competent, warm and trustworthy you look to people who have never met you.

That variation isn't cosmetic — it's measurable. In a 2011 study in Cognition, Jenkins, White, Van Montfort and Burton gathered photos of the same faces from the open web and asked unfamiliar viewers to sort them by identity. Viewers who didn't know the people routinely split one person's photos into multiple identities, while viewers who did know them sorted perfectly. The researchers concluded that within-person variability — how much you differ from photo to photo — can exceed between-person variability. In other words, the gap between your best and worst photo can be larger than the gap between you and a stranger.

And the judgment happens fast. In a well-cited 2006 study, Janine Willis and Alexander Todorov showed faces for just 100 milliseconds and found the resulting trait judgments — trustworthiness, competence, likeability — correlated highly with judgments made with unlimited viewing time. Extra time mostly increased people's confidence in the impression they had already formed. Your headline, your bio and your credentials do not get a chance to argue with that first tenth of a second.

AI generation raises the stakes because it produces divergence rather than convergence. A photographer shooting you for twenty minutes gets fifty variations of one look. A generator gets fifty attempts at you — and some of those attempts are a slightly different person: a narrower jaw, symmetrical eyes, five years younger, teeth you don't have. Plausible is not the same as correct, which is why likeness has to be the first filter, not the last.

Start with a two-pass shortlist, not a deep analysis

Do not analyze fifty images. Fatigue sets in around image fifteen and your standards quietly collapse. Instead, run two fast passes: an instinct pass that cuts the batch by two-thirds, then a structured pass on what survives. The whole shortlist should take five minutes, not an evening of second-guessing.

1. 1Pass one (instinct, 3 seconds each): keep or kill, no reasoning allowed. If you hesitate, kill it. Aim to land at 12–15 keepers.
2. 2Deduplicate: near-identical shots (same pose, same background, same expression) count as one. Keep the sharpest, discard the twins.
3. 3Pass two (structured): open the survivors large — full screen, not a thumbnail grid — and apply the five tests below.
4. 4Compare in pairs, never in a crowd: put two images side by side and eliminate one. Repeat. Tournament brackets beat staring at a wall of faces.
5. 5Check them small too: shrink your finalists to roughly the size of a comment-thread avatar. Most profile photos are seen tiny, so the tiny version is the real product.
6. 6Stop at three finalists: more than three and you are no longer choosing, you are shopping.

One process note that saves people a lot of pain: decide what the photo is for before you open the folder. A photo chosen for a law-firm bio page and a photo chosen for a creator's newsletter are different photos, and a shortlist built without a target audience tends to converge on whatever is most flattering rather than whatever is most useful.

Test 1: likeness — does it look like you today?

Likeness is a pass/fail gate, not a scoring dimension. If a shot doesn't read as the current version of you, no amount of beautiful lighting rescues it — the person who meets you after seeing it experiences a small correction, and small corrections cost trust. Run this test before you let yourself notice which images are pretty.

The most useful trick is a direct reference comparison. Open a recent, honest photo of yourself — a phone snapshot from the last few months, not your favourite photo from 2021 — and put it next to each candidate. Then compare specific landmarks rather than overall vibe: hairline shape, jaw width at the ears, nose bridge, eye spacing, the asymmetries you actually have. Vibe is easy to fake; geometry isn't.

• Age and skin: generators tend to smooth and de-age. If your laugh lines and greys have vanished, that's drift, not retouching.
• Weight and face shape: a slimmed jaw or narrowed cheeks is the single most common flattering error.
• Symmetry: real faces are asymmetric. An image that has quietly mirrored you into balance will feel uncanny to people who know you.
• Hair: current length, current density, current hairline. A fuller version of your hair is a tell.
• Glasses and facial hair: they should match what you wear to meetings, not what the model preferred to draw.
• Teeth: straighter, whiter, more numerous teeth are a generation artifact more often than a win.

Be honest about the input problem, too. A generator learns you from the selfies you uploaded, so if your inputs were mostly three years old, mostly one angle, or mostly the same lighting, the output will reflect that narrow view. When several finalists fail likeness in the same direction — all younger, all thinner, all differently-haired — the fix is upstream, in the input set, not in the picking.

Test 2: expression and approachability

Once likeness passes, expression is the highest-leverage variable left. It's also where people most often pick badly, because we judge our own expressions against how we felt, not how we look. The eyes decide most of it: engaged eyes with a little lower-lid tension read as confident, while wide, slack eyes read as startled.

A genuine smile involves the eyes as well as the mouth — the crinkling at the outer corners is what separates warm from posed. Photofeeler, which crowdsources ratings of profile photos, analysed roughly 60,000 ratings across 800 business profile photos and reported that a teeth-showing smile was the largest single positive driver in the set: around +1.35 on perceived likability, with smaller gains of about +0.33 on competence and +0.22 on influence. A closed-mouth smile delivered roughly half the likability benefit and no meaningful competence or influence gain. In the same analysis, the slight narrowing of the lower lids that photographers call a "squinch" was associated with gains of about +0.33 on competence and +0.37 on influence.

That's one company's proprietary dataset, so treat the exact numbers as directional rather than as a law. But the direction is corroborated in an unusual place: LinkedIn's own photo guidance tells members that people view you as more likable, competent and influential when you smile, and that smiles showing teeth were rated roughly twice as likable as closed-mouth smiles. Two independent sources pointing the same way is about as much certainty as this subject offers.

Calibrate to the room you're trying to enter:

• Sales, recruiting, consulting, real estate, anything client-facing: open, warm, teeth-visible smile. You are selling approachability.
• Law, finance, medicine, security, government: composed, closed-mouth or lightly smiling. You are selling steadiness.
• Founders, creatives, marketers, speakers: more personality is allowed and often rewarded — a real laugh, an angled head, a little asymmetry.
• Academia and research: neutral and unfussy; over-produced expressions read as marketing.

Expressions to discard regardless of industry:

• Smiling mouth with dead, unengaged eyes — the most common AI expression failure.
• One mouth corner noticeably higher than the other, which reads as a smirk at small sizes.
• Raised eyebrows plus wide eyes, which reads as surprised or anxious.
• An over-stretched grin that pulls the cheeks flat and flattens the eyes.
• Anything you would not do in an actual first meeting. If you wouldn't hold that face for three seconds in person, it's the wrong photo.

Test 3: framing, crop and surviving the circle

A great expression can still be undone by geometry. Profile photos are displayed as small circles across nearly every platform, so the frame has to hold up after the corners are cut and the image is scaled down to a fraction of its size. Judge framing at final display size, not at 100% on a big monitor.

LinkedIn's own guidance is unusually specific about this: aim to have your face filling about 60% of the frame, and crop from the top of your shoulders to just above your head. There's a floor as well as a ceiling, though. Photofeeler's data found face-only close-ups pulled likability down slightly (about −0.21), while full-body shots hurt perceived competence and influence. Head-and-shoulders, or head-to-waist, is the range that performs.

• Crop: head-and-shoulders — roughly crown of the head to mid-chest — is the most reliable range for profile use.
• Eye line: eyes sitting in the upper third of the frame reads naturally; eyes dead-centre reads like an ID photo.
• Face size: around 60% of the frame, per LinkedIn's guidance, keeps you recognizable when the image shrinks.
• Headroom: a sliver above the hair, not a canyon. Excess empty space above the head shrinks your face in the circle.
• Edges: chin, ears and shoulders should not be clipped by the frame — the circular crop will bite in further.
• Orientation: square or comfortably croppable to square. A wide landscape shot loses too much.

The technical floor is easy to check. LinkedIn accepts a PNG or JPG profile photo between 400 × 400 and 7680 × 4320 pixels at up to 8 MB, and crops it to a circle — displayed large on your profile page, but only a few dozen pixels across in the feed and smaller still in comment threads. Slack's help documentation asks for a square image no smaller than 512 × 512 and no larger than 1024 × 1024. Microsoft 365 stores profile photos in a fixed ladder of sizes that runs all the way down to 48 × 48, which tells you how small your face can get in an Outlook or Teams interface. Our reference guide to LinkedIn photo sizes and specs has the full numbers; for selection purposes, the practical test is simpler: crop your finalist to a circle, scale it to roughly 48 px, and see whether it still reads as a confident human being.

Test 4: background and wardrobe fit for your industry

Backgrounds are the easiest place to accidentally send the wrong signal, and the easiest place to spot generation errors. The rule is fit, not preference: the background should look like the environment your audience associates with competence in your field, and it should be quiet enough that your face wins the attention contest.

• Seamless grey, white or soft gradient: corporate, finance, legal, medical, enterprise sales. Safest and most timeless.
• Blurred office or interior: consulting, real estate, agency, mid-market B2B. Adds context without clutter.
• Outdoor or natural light: wellness, nonprofit, education, creative and coaching work.
• Dark, low-key background: speakers, executives, technical founders — high contrast, high drama, less versatile.

Discard any background where architecture bends, signage turns into pseudo-text, a doorway or shelf line slices through your head, or a plant appears to grow out of your skull. Also watch luminance: dark hair against a dark background loses the outline of your head, which is exactly what makes an avatar unreadable at 48 px. Our guide to choosing a headshot background covers the trade-offs in more depth.

If your batch produced different outfits, apply the same fit logic. Pick the one you would genuinely wear to meet this audience rather than the most expensive-looking one — an outfit that outranks your actual working wardrobe creates the same small correction that a wrong likeness does. Photofeeler's dataset found formal dress was the strongest single driver of perceived competence and influence in their sample, so if your field rewards those traits over warmth, the more formal finalist is usually the right call.

Test 5: technical quality — where AI images still break

Now zoom in. Modern generators fail less often than they did two years ago, but when they fail they fail strangely, and the failures cluster in predictable places. Inspect each finalist at 100% and walk the checklist in order. One artifact on a focal feature — eyes, teeth, glasses — is a discard, not a fix-it-later.

A peer-reviewed guide to distinguishing AI images from real photographs, by Kamali and colleagues, sorts the giveaways into five families: anatomical implausibilities, stylistic artifacts, functional implausibilities, violations of physics, and sociocultural implausibilities. For headshots, the first four do nearly all the work — and they map neatly onto a checklist you can run in two minutes.

1. 1Eyes: catchlights should sit in roughly the same position in both eyes, pupils should match in size, irises should be round and consistently coloured.
2. 2Teeth: look for individual teeth with visible separations and slight irregularity, not a smooth white band or an implausible extra row.
3. 3Glasses: arms should reach the ears, the lens edge should distort what's behind it, and the frame should cast a small shadow rather than sit on the face like a sticker.
4. 4Ears and earrings: paired items should match without being identical. Two different earrings is the classic giveaway.
5. 5Hair edges: individual strands at the outline, not a soft halo where hair dissolves into the background.
6. 6Collars, lapels and buttons: both sides should follow the same geometry; melted plackets and vanishing buttonholes are common.
7. 7Skin: visible pores and texture. Plastic, poreless skin is one of the most frequently cited AI tells, and the easiest to spot at a glance.
8. 8Lighting physics: shadows should fall in one consistent direction, and the light on your face should match the light in the background.

Text deserves a special note, because the old advice has expired. Generated text on lanyards, badges, signage or book spines used to be reliably garbled, and "look for gibberish" was the single fastest AI test available. The best 2026 image models render short text correctly a great deal of the time, so read the text rather than assuming it's wrong. Garbled wording is still a hard discard; clean-looking text on a badge or credential you don't actually hold is a different problem, and worth cropping out either way.

Then reverse the zoom. Some micro-artifacts are invisible at display size and genuinely don't matter; others are subtle at 100% but scream at thumbnail size because they distort the overall shape of the head. Judge twice — once at 100% for artifacts, once at avatar size for impression — and only keep images that pass both. We break the failure modes down further in 10 signs an AI headshot looks AI-generated.

Get a second opinion — your own judgment is biased

This isn't a courtesy step; it's a correction for a documented bias. Across two internet-based studies with 610 participants, White, Sutherland and Burton found that people made suboptimal choices when selecting profile images of their own face: images chosen by strangers produced more favourable impressions than the ones people picked of themselves.

The detail that should worry you is where the penalty landed. The self-selection disadvantage showed up on perceived trustworthiness and competence rather than attractiveness — meaning you can confidently pick the photo you find most flattering and still pick the one that costs you credibility. (A 2021 correction to the paper fixed a data-processing error in one calibration analysis; the authors reported the main effects held and were in fact strengthened.)

The mechanism is simple. You have decades of experience seeing your own face in mirrors and in photos you liked, so you evaluate candidates against an internal self-image rather than against the impression a stranger receives. Other people have no such baggage. They just see a face and react — which is precisely the reaction you are trying to optimize.

• Recruit two groups: one or two people who know you well (they judge likeness) and two or three who don't (they judge impression).
• Show three to five finalists, unlabelled, in a random order, with no commentary from you.
• Ask one question per group: "Which of these looks most like me?" for the people who know you, and "Which of these would you take a meeting with?" for the ones who don't.
• Ask for a ranking, not a rating. Forced choices produce signal; five-star scores produce politeness.
• Never reveal your own favourite before they answer, and don't defend it after.
• Count votes, ignore commentary. Aesthetic notes are noise; the tally is the data.

Pick per channel: LinkedIn vs. resume vs. email vs. Slack

Different surfaces render your face at different sizes, in different shapes, to different audiences. Your finalists usually split cleanly across them — one warm shot, one conservative shot, one tight crop — so decide deliberately instead of uploading the same file everywhere by default and hoping it works.

• LinkedIn: the highest-stakes slot. Warm expression, direct eye contact, head-and-shoulders crop, quiet background, and proven readable at feed and comment sizes.
• Resume or CV: only where it's customary — norms differ sharply by country, as our country-by-country rules on resume photos lays out. Where a photo is expected, choose the most conservative finalist: neutral background, composed expression.
• Email signature: often rendered at 60–100 px and frequently on a white background. Pick the highest-contrast, simplest, tightest crop you have.
• Slack and Teams: seen daily by colleagues, so approachability beats gravitas. A friendlier, slightly more relaxed shot works better here than your most formal option.
• Company team page: consistency with teammates matters more than personal preference — matching crop, matching background family, matching light direction.
• Speaking bios and press: more environment, more personality, higher resolution. Event organisers often crop and print, so give them room.

One caveat that cuts against variety: recognition compounds. If a prospect sees you on LinkedIn, then in Slack Connect, then in a meeting invite, the same face in all three places is worth more than three slightly better-optimized faces. Use one primary image across the professional channels you care about, and only diverge where the medium genuinely demands it.

What to discard — and when to run it again

Most of a batch should end up deleted, and that's the system working as intended. Generation is cheap, publishing is not. Being decisive about discards is what separates a headshot you're happy to be recognized from and a headshot you quietly replace in three weeks.

• Fails likeness in any way you'd have to explain in person.
• Any artifact on a focal feature — eyes, teeth, glasses, ears.
• An expression you would not repeat in a real first meeting.
• A crop that loses your chin, ears or shoulders once it's circular.
• A background that confuses, dates or camouflages you.

Resist the sunk-cost pull. Keeping a "nearly right" shot because you paid for the batch is the most expensive small decision in this whole process. If none of your candidates survive the likeness gate, the problem is almost always the input set — too few angles, too old, too much of the same lighting, heavy filters, sunglasses, group shots. A better input set usually changes the output more than any prompt tweak, and our guide to input selfies that actually work covers what to shoot before you run a new batch.

And sometimes the honest answer is that generation is the wrong tool for this job. Full-body shots, real environments, product-in-hand images, a specific brand set, or a look you need to art-direct in real time all favour a photographer, who can also take direction on the spot, adjust the light for your particular face, and guarantee that the person in the frame is unambiguously you. Generators also drift more often with very distinctive features — pronounced asymmetry, unusual hair, head coverings, visible medical devices — simply because there is less of that in the training distribution. If you're weighing the two paths, our honest comparison of AI headshots and photographers lays out where each one wins.

The 10-minute selection workflow

Put together, the whole process is a single short session. Do it once, properly, and you won't reopen the folder every month. Set a timer, work in order, and don't let yourself skip back to admiring images you've already eliminated.

1. 10:00 — Write down the primary use (LinkedIn, team page, speaker bio) and the audience.
2. 20:30 — Instinct pass, 3 seconds per image, down to 12–15 keepers.
3. 32:00 — Deduplicate near-identical shots.
4. 42:30 — Likeness gate against a recent real photo. Anything that fails is out.
5. 54:00 — Expression pass: eyes first, then mouth, calibrated to your industry.
6. 65:30 — Framing and circle test at 48 px; background and wardrobe fit.
7. 77:00 — Artifact inspection at 100% on the last three or four.
8. 88:00 — Send three finalists to your panel, ranked answers only, then upload the winner and file the runner-up for Slack and email.

The goal was never to find the best-looking image in the batch. It was to find the image that a stranger reads correctly in a tenth of a second and that you can still walk into a meeting behind. Those are frequently not the same photo — and knowing the difference is the entire skill.