AI Headshots and Natural Hair, Hijab and Beards: Where the Models Still Fail

AI headshots and natural hair are still an uneasy match. Image models learn from what the internet contains, so coils, locs, braids, hijab, full beards, tattoos and darker skin tones break more often than a clean-shaven face under studio light. Here is why, what to check in your own results, and when to book a photographer instead.

We build an AI headshot product, which makes this an awkward article to write and a necessary one. The same system that returns a usable portrait in ten minutes for one person returns an unusable one for another, and the pattern of who gets failed is not random. It tracks the shape of the training data with depressing consistency. Everything below is either documented in published research or checkable in your own results in about ten minutes.

Why these failure modes exist at all

None of this is deliberate. It falls out of two things: what was in the training data, and what the model learned to associate with the word professional. Once you hold both in mind, the specific breakages stop looking like random bugs and start looking like predictable consequences of how these systems are built.

The training set is a snapshot of the internet, not of the world

LAION-5B, the open dataset behind several widely used image models, contains 5.85 billion image-text pairs, of which 2.32 billion are English. Representation in a web-scraped set follows whoever posts, whoever gets photographed by professional outlets, and whoever gets captioned in English. That is not a neutral sample of humanity, and no amount of clever architecture fixes a sample problem.

The lineage is old. In 2018, Joy Buolamwini and Timnit Gebru's Gender Shades study found that three commercial gender-classification systems misclassified darker-skinned women at rates as high as 34.7%, while the worst error rate for lighter-skinned men was 0.8% — largely because the standard benchmark sets were overwhelmingly composed of lighter-skinned subjects. Classification is a different task from generation, but the underlying data economics are the same.

A 2025 experimental study in the Journal of the European Academy of Dermatology and Venereology put numbers on the generation side. Across 4,000 AI-generated clinical images from four generators, only 10.2% depicted dark skin — 6.0% for ChatGPT-4o, 3.9% for Midjourney, 8.7% for Stable Diffusion. Adobe Firefly produced 38.1%, the closest alignment with US demographic data of the four. That last number matters more than the others: it shows the skew is a design choice, not a law of physics.

The model carries a prior about what 'professional' looks like

Bloomberg's 2023 analysis generated more than 5,000 Stable Diffusion images — 300 each for seven higher-paying and seven lower-paying occupations, plus three crime-related categories — and found higher-paid roles such as CEO, lawyer and politician dominated by lighter skin tones, while lower-paid roles such as fast-food worker and dishwasher skewed darker. The same analysis produced almost three times as many images of men as of women. The model was not asked to do this. It learned it from captions.

The most-cited consumer example came in July 2023, when Rona Wang, then an MIT student, uploaded her photo to Playground AI and asked for a professional LinkedIn profile photo. It returned a version of her with lighter skin, lighter hair and blue eyes. Playground's founder Suhail Doshi replied publicly that the models "aren't instructable like that" and will pick a generic result based on the prompt. That is a fair technical description of the mechanism, and it is exactly why the burden of checking falls on you.

A 2025 Scientific Reports paper by AlDahoul, Rahwan and Zaki examined Stable Diffusion across six races, two genders, 32 professions and eight attributes, and documented what the authors call racial homogenization — nearly all Middle Eastern men rendered as bearded, brown-skinned and in traditional attire. The same paper ran a preregistered survey experiment finding that exposure to non-inclusive AI-generated faces increased participants' own racial and gender biases, while inclusive ones reduced them.

Corrective fixes are hard to aim

In February 2024, Google paused Gemini's ability to generate images of people after its diversity correction produced historically nonsensical outputs. It is worth remembering that episode when a vendor promises they have solved bias: crude post-hoc balancing breaks in different, equally visible ways. The durable fixes are unglamorous — better dataset composition, and evaluation against a scale that actually resolves darker skin, such as the ten-shade Monk Skin Tone Scale that Google open-sourced in 2022 with Harvard sociologist Ellis Monk, precisely because the older Fitzpatrick scale under-represents dark shades. If you want the mechanics of how generators are trained on your uploads in the first place, we cover that in how AI headshot generators actually work.

AI headshots and natural hair: coils, locs, braids and the edge problem

Textured hair is the single most common failure in consumer AI headshots, and it fails in a specific way: the model produces something hair-shaped and plausible at thumbnail size that falls apart the moment you zoom. The pattern is structural, because hair is the highest-frequency detail in a portrait and among the least represented in training data.

Here is what actually breaks, in rough order of how often you will see it:

• Coil pattern flattened. 3C–4C texture gets rendered as a generic soft wave or a uniform digital curl, losing the definition that makes the style recognisably yours.
• Locs rendered as identical cylinders. Mature locs vary in diameter, direction and root tension. Models tend to produce evenly spaced tubes of equal thickness, which reads as plastic.
• Invented parts. Cornrow rows drift, merge or terminate in the middle of the scalp. Braid parts appear where you never had one. Feed-in patterns lose their direction halfway back.
• Hairline errors. Baby hairs, a receding hairline, a widow's peak or wig lace all sit at the hardest boundary in the image and are frequently smoothed away or redrawn in the wrong place.
• Floating volume. An afro or a puff that does not connect convincingly to the scalp — the silhouette is right, the attachment is wrong.
• Inconsistency across the batch. Length, shrinkage and style vary between outputs from the same upload set, so no two images look like the same person on the same day.

The mitigation that moves the needle most is upstream, in what you upload. Give the model many angles of your current style — not a mix of last year's twist-out and this month's box braids — with even light reaching the top and back of the head, so the shape is unambiguous. Our guide to input selfies for AI headshots covers the full upload spec; the short version for textured hair is more angles, one style, no hats.

Hijab and other head coverings

This one has been documented directly, on this exact product category, and the finding is blunt. In research published by UC Berkeley Law in March 2026, J.S.D. candidate Mahwish Moazzam reported testing around 25 widely used AI headshot generators over roughly a year, uploading her own photos wearing hijab, and retesting several of them months later.

The apps removed the hijab and replaced it with generated hair. Only two returned anything resembling a covering, and those were distorted or incomplete. Moazzam noted that several of the apps explicitly asked whether she wanted to keep accessories such as glasses. None asked about the hijab. Her argument is that this is not a rendering glitch but a systematic alteration of a visible religious identity marker, executed without consent and at scale — and that anti-discrimination law, built around identifiable human decision-makers, is poorly equipped to assign responsibility when the decision-maker is an algorithm.

The mechanism is the same one behind every other item on this page. A model trained mostly on uncovered heads treats a covering as an obstruction between it and the face it expects to find. The same logic applies to a Sikh dastaar or patka, a kippah, a sheitel, a headwrap, a kufi, and head coverings worn during chemotherapy. If your covering is non-negotiable — and for most people who wear one, it is — treat every generator as guilty until proven otherwise.

What to check before you spend anything, on every image in the batch rather than the two the interface shows you first:

• Is the covering present in all outputs, or does it disappear in a subset?
• Does hair leak out at the temples, the nape or the forehead where it should not?
• Is the drape line under the chin continuous, or does it dissolve into the neck or the collar?
• Are pins, folds and the ear region rendered as fabric, or as a smeared texture that is neither hair nor cloth?
• Is the fabric colour and opacity consistent across the set, or does it shift between images?

Accessories generally are the weak point of this technology. The same instability that eats a headscarf also produces warped frames and phantom reflections on eyewear, which we unpack in our guide to wearing glasses in headshots.

Full beards, facial hair and religious grooming

Facial hair fails in both directions, which is what makes it interesting. Models both under-render the beards people actually have and over-assign beards to people who do not have them — the homogenization effect the Scientific Reports authors measured, where an entire demographic gets collapsed into one bearded template.

On the under-rendering side, the recurring problems are density and boundary. Sparse or patchy growth gets filled in denser than reality, because dense beards dominate the training images. The neckline — where you actually trim — is invented rather than observed, so a carefully maintained line becomes a soft fade or a hard edge you never wore. The mustache-to-lip transition turns mushy at full resolution. Gray distribution shifts, usually toward less gray. And long beards worn for religious reasons, whether Sikh, Muslim or Orthodox Jewish, are frequently shortened or tidied into a corporate template, which is the grooming equivalent of the hijab problem.

Three checks catch most of it:

1. 1Zoom to 100% on the neckline and the cheek line. If either boundary is a shape you have never worn, the model invented it.
2. 2Compare gray coverage against an unedited recent selfie. Silent de-aging is common and it is the detail people notice when they meet you.
3. 3Check length consistency across the batch. If the beard is three different lengths across 50 outputs, the model is guessing rather than reproducing.

Visible tattoos

Tattoos on the neck, hands or forearms sit in an unhappy middle ground. A generator has no representation of your specific artwork; it has a general notion of what ink on skin looks like. So it does one of three things, and all three are problems if the tattoo matters to you.

• Erases it. Many pipelines effectively treat unusual skin markings as noise to be cleaned, alongside blemishes and stray hairs.
• Approximates it. You get a tattoo in roughly the right place with the wrong content — which reads worse than no tattoo, because anyone who knows you can see it is not the same piece.
• Garbles the lettering. Small text is a well-known weak point of image models. Script tattoos come back as convincing-looking nonsense.

Your realistic options are to keep the ink out of frame with a tighter crop, to accept a clean-skin portrait and be comfortable that your headshot omits something visible about you, or to photograph it properly. If your tattoos are part of how clients and colleagues recognise you — common in creative, hospitality, fitness and trades work — an AI portrait that quietly removes them is not a flattering version of your photo, it is a photo of somebody else.

Darker skin tones and skin-tone shifting

Skin-tone drift is the failure with the longest paper trail, and it predates generative AI by a decade. It shows up in three distinct forms: overall lightening, undertone shift, and feature drift — where the model does not just brighten the skin but nudges nose, lip and eye shape toward a different template.

The Rona Wang case is the vivid consumer example, but the systematic evidence sits in the filter research. MIT Technology Review documented in 2021 how beauty filters perpetuate colorism, and a 2024 peer-reviewed study by Riccio, Colin, Ogolla and Oliver in Social Media + Society — bluntly titled Mirror, Mirror on the Wall, Who Is the Whitest of All? — found that social media beauty filters embed Eurocentric beauty canons not only by brightening skin but by modifying facial features, so that beautified faces are more likely than the originals to be classified as White. Generative headshot tools inherit the same aesthetic priors from the same web imagery.

Auditing this on your own results takes about two minutes and requires one thing: an unedited reference selfie taken in neutral, indirect daylight, open side by side with the generated image at the same size on the same screen.

1. 1Compare undertone, not just brightness. A result can be the same lightness value and still shift warm-to-cool or olive-to-pink, which is the change people register as 'that doesn't look like me'.
2. 2Look at specular highlights. On deep skin, over-brightened highlights read as ashy or plastic rather than luminous — a common tell that the model applied a lighting template built for lighter skin.
3. 3Check that only tone changed. Trace the nose bridge, lip fullness and eye shape against your reference. If the geometry moved, the model substituted a face rather than lit yours.
4. 4Check the batch, not the favourite. Tone often drifts progressively across a set of 50, and the images the interface surfaces first are rarely the representative ones.

There is an important distinction here between bias and ordinary retouching, and it is worth keeping straight: evening out a blemish is retouching, changing your complexion is not. Our piece on how much headshot retouching is too much sets the line for the first case, and the ten signs an AI headshot looks AI-generated covers the artefacts that make a result unusable regardless of tone.

Age and visible disability

Two more categories where the research is clear and the consumer impact is under-discussed. Both share a common shape: the model has a thin, stereotyped representation of the category, so it either erases the trait entirely or exaggerates it into a cliché. Neither outcome produces a portrait that looks like the person who sat down to make it.

Age

Multiple studies have documented what researchers call digital ageism in image generators. A longitudinal comparison published in the Journal of Medical Internet Research in 2025 regenerated 164 DALL-E 2 images from identical geriatric-lexicon prompts a year apart and found the patterns essentially unchanged, with White-racialized older adults depicted about five times more often than all other groups in both years. A separate visual-properties analysis of 456 Midjourney images found that pictures of older people were significantly less bright and less sharp than pictures of younger people, with more smiling and more glasses.

For headshots the practical harm is the opposite of stereotyping: silent de-aging. Generators routinely remove gray, soften nasolabial folds and tighten the jawline without being asked. That is a problem on its own terms, and it is a problem on interview day, when the person who arrives does not match the profile. We cover the career side of this in age bias, your LinkedIn photo and what you actually control, which deals with the hiring dynamics rather than the rendering.

Visible disability

The reference work here is a 2024 CHI paper by Kelly Avery Mack and colleagues, based on eight focus groups with 25 people with a range of disabilities. Its title is the finding: They only care to show us the wheelchair. Participants described models that repeatedly fell back on reductive archetypes — over-associating disability with wheelchair use, depicting disabled people as sad or in need of help, and reducing blindness to a pair of dark glasses.

In a headshot the failure mode is narrower but just as consequential. Hearing aids, cochlear implant coils, cannulas, ptosis, facial differences, vitiligo and craniofacial variation all tend to be treated as artefacts to be cleaned up, because the model's prior for a professional portrait does not contain them. If a visible trait is part of how you are recognised, check specifically for its presence in every output, and treat its absence as a failed batch rather than a flattering result.

Non-Western formal wear

If you want to be photographed in a sherwani, an agbada, an ao dai, a barong, a hanbok or a sari, expect the model to be markedly less reliable than it is with a navy blazer — and expect the unreliability to run in two opposite directions at once.

The most careful work on this is a community-centred study of text-to-image models in South Asia, published at the ACM FAccT conference in 2023, in which 36 participants from Pakistan, India and Bangladesh co-designed prompts and reviewed the outputs together. The researchers sorted what they saw into three failure modes: the models fail to generate culturally specific subjects at all, they fall back on hegemonic cultural defaults, and they perpetuate cultural tropes. The authors describe the result as reproducing an outsider's gaze on the cultures depicted.

For a portrait, those last two failure modes are the ones that bite, and they pull against each other. The default pulls your outfit toward whatever the model associates most strongly with your region or name, whether or not you asked for it; and when you do ask for a specific garment, its understanding of the construction is shallow enough that the details come back approximated. Practically, that means naming the garment precisely rather than gesturing at a region, uploading input selfies in which you are actually wearing it, and auditing the specifics — pleat direction, collar structure, drape over the shoulder, where jewellery sits. The further your outfit sits from a Western business template, the more likely the model is to invent instead of reproduce.

A ten-minute audit of your own results

Whatever tool you use, including ours, the same audit applies. Do it before you pay for downloads, not after, and do it on the full set rather than the handful the interface promotes. Almost every failure above is visible in under ten minutes if you know where to look.

1. 1Open an unedited reference selfie, shot in neutral indirect daylight, next to the results at identical size.
2. 2View at 100%, not at thumbnail size. Thumbnails hide every artefact discussed on this page.
3. 3Check the identity markers first — hair texture and style, covering, beard, tattoos, hearing aid, glasses. Present and correct, or absent?
4. 4Trace the boundaries: hair against background, fabric against neck, beard against jaw, glasses against temple. Boundaries are where generators break.
5. 5Compare tone and undertone against your reference, then compare feature geometry separately.
6. 6Look for de-aging: gray coverage, skin texture, jawline. Ask whether this is you on a good day or someone five years younger.
7. 7Scan the whole batch for consistency. If the same upload set produces three different hair lengths or two different skin tones, the model is generating rather than reproducing.
8. 8Apply the recognition test: would a colleague who has met you twice identify this as you across a meeting room?

If a result fails on an identity marker, no amount of good lighting redeems it. Discard it rather than talking yourself into it — and if the whole batch fails the same way, that is information about the tool, not about you.

When a human photographer is simply the better call

We would rather tell you this than sell you a batch you cannot use. There are cases where the current generation of this technology is the wrong tool, and recognising yours early saves you a fortnight of regenerating and squinting at outputs that were never going to converge.

• You wear a hijab, turban, patka or other religious covering and it is non-negotiable. Given the Berkeley findings, treat AI as a free experiment at best and plan around a photographer.
• You have an intricate protective style — knotless braids, a specific loc pattern, a fade with a design — that you need reproduced exactly rather than approximated.
• Your visible tattoos are part of your professional identity.
• You have a facial difference, a visible disability or an assistive device you want represented accurately and with dignity, rather than cleaned away.
• Your photo carries legal or likeness weight — real estate signage, bar association profiles, medical directories, expert-witness work — where an inaccurate portrait is a liability rather than an aesthetic choice.
• You have already run one batch and the same identity marker failed across all of it. A second batch rarely fixes a structural gap.

None of that makes photographers universally superior. They cost meaningfully more, they take a week or more of scheduling, and a mediocre photographer produces a worse result than a good generator. We laid out that trade in detail in AI headshots versus a photographer, which weighs cost, turnaround and control across the general case; this article is the narrower argument that the decision should also turn on whether your face and hair sit inside or outside the model's competence, not on price alone.

What to ask any AI headshot vendor before you pay

Bias claims are unfalsifiable in marketing copy and trivially checkable in product. These are the questions that separate a vendor who has thought about this from one who has not, and every single one is answerable before you hand over money.

• Can I see the results before paying? Anything that charges upfront for an unseen batch is asking you to absorb their failure rate.
• Can I inspect at full resolution? Thumbnail previews conceal precisely the artefacts described above.
• Do the published samples include people who look like me? A gallery of one demographic is a statement about the training and evaluation behind it.
• Is there any commitment to preserving identity markers? Very few vendors make one. Notice which do.
• What happens to my uploads? Ask for the retention and deletion policy in writing before you upload anything.
• Is the output disclosed as AI-generated? A vendor that is straightforward about what it produces is more likely to be straightforward about what it cannot do. Ours is set out on our AI disclosure page.

Where this leaves you

The honest summary is that AI headshots work well for a face the training data saw a great deal of, and progressively less well as you move away from that centre. Firefly's 38% figure in the dermatology study and the existence of the Monk Skin Tone Scale both point the same way: this is fixable with deliberate dataset and evaluation work, and it will get better. It is not fixed today.

Until it is, the leverage sits with you and it is entirely mechanical. Upload inputs that show your actual hair, covering and grooming from many angles. Audit at 100% against an unedited reference. Refuse any result that erased something real about you, however flattering the lighting. And when the failure repeats across a whole batch, stop paying tools to approximate you and book someone with a camera.