What AI hair analysis can and cannot see in a photograph

Think about what is physically present in the file. Colour, brightness, edges, shadow, and the shapes those make. A model reads patterns in that, compares them against the many thousands of hair photographs it has seen, and describes what your hair most resembles.

That means anything visible at the scale of a whole head is available, and anything that needs magnification, touch, or a chemical test is not. It is a useful dividing line, and you can apply it yourself to any claim a tool makes. Ask whether the thing being claimed leaves a mark in the light. Curl shape does. Water absorption does not.

The good news is that a lot of what decides your routine is visible. The bad news is that three of the readings people most want are on the other side of the line.

Curl pattern is the strongest reading, because it is a shape and shapes are what photographs are made of. The size and repeat of the bends, whether the curl holds at the ends or drops, and where it loosens across the head all show plainly in daylight on dry hair.

Visible density at the part and the crown is the second. How much scalp shows through along a part line is exactly the kind of light-and-dark pattern a photograph carries well. So is a regrowth line, where a band of different colour or texture meets the treated hair. That band is a date stamp: hair grows about 0.35 mm a day, roughly 1 cm a month per StatPearls, so a 3 cm band puts the last colour appointment around three months ago, and the model can read that without being told.

The rest of the visible list is condition. Rough or split ends catch light differently from smooth ones and read as a frayed, dull edge. Shine at the roots with a duller length says oily scalp with dry mid-lengths. Colour condition shows in the difference between the root colour and the ends. And the hairline is clear in a straight-on photograph: its shape, whether it has moved back at the temples, and how fine the hairs along it are.

Stand in front of a bathroom mirror and you see your face and the front of your hair. You do not see the top of your own head, and you almost never see the crown, which is where thinning and product build-up show earliest.

That is the single strongest argument for a photographic reading over a self-assessment. A camera held above your head, or a phone handed to someone else, produces the one view you have been missing for your entire life. People are often surprised by it, and the surprise is not always bad news: plenty of crowns look denser in the photograph than their owner expected, because the mirror only ever showed them the part line at the front.

It is also the photograph most people skip when a tool asks for three. If you take only one extra, take that one.

Three readings sit permanently out of reach, and any tool that claims them from an ordinary photo is inferring rather than measuring.

Strand width is the first. The difference between a fine strand and a coarse one is measured in tens of microns, and a phone photograph of a whole head has nowhere near that resolution. A model can guess from how the hair behaves as a whole, and the guess is often reasonable, but it is a guess. The reliable version takes two seconds with your own fingers, and it is described at hair density and thickness.

Porosity is the second, and it is not visible at any resolution, because it is about how the cuticle takes up and holds water. You test it by wetting hair and watching what happens, not by looking at it dry. Elasticity is the third, and it is a mechanical property: you find it by stretching a wet strand and seeing whether it returns. A photograph cannot pull anything. The tests for both are at hair porosity, and they take longer to read about than to do.

This sounds obvious and it is the most common reason a reading is wrong. The back of the head does not exist unless it was photographed. Neither does the underside, the nape, or the section behind the ear where many people carry the most breakage.

Hair is rarely uniform. The nape is often finer than the crown, curl pattern can loosen at the front, and damage collects where a hair tie sits or where a straightener is easiest to reach. A reading from a front-facing photograph describes the front of your head and then, if it is careless, states it as a fact about all of you.

A photograph also cannot smell. That matters more than it sounds: a scalp that smells by the second day is a real signal about oil and bacteria, and it is one of the few things you can assess better than any tool can. Texture under the fingers, tenderness when you press, and how the hair feels when it is wet are all yours alone.

This is a hard boundary rather than a limitation to be improved on later. Flaking, redness and itch look similar in a photograph across conditions that need different treatment. The American Academy of Dermatology names several that sit behind what looks like ordinary dandruff: seborrheic dermatitis, psoriasis, fungal infections of the scalp, and eczema.

Look at what a dermatologist does, and the gap is obvious. The American Academy of Dermatology describes the visit as asking how long the loss has been happening and whether it came on quickly, examining the scalp and nails and any other area with hair loss, and gently pulling the hair to see how it is growing. If the cause could be a disease, a vitamin deficiency, a hormone imbalance or an infection, the next step is a blood test or a scalp biopsy. None of those five things is available to a photograph.

So the correct output from an AI reading, when it sees something that might be medical, is a sentence telling you to go and get looked at. What that looks like in practice is at when a hair consultation should send you to a doctor.

There is published work on machines counting hair, and the conditions it needs are a useful reality check on what a selfie can do.

A 2022 study in Sensors evaluated automated hair density measurement using deep neural networks. The dataset was 4492 enlarged scalp images taken from 817 male hair-loss patients, every image at 1280 by 1024 pixels, close enough to resolve individual follicles and count how many hairs came out of each one. The best of the three models tested, YOLOv4, reached 75.73% accuracy with a precision of 80.75%.

Read that twice. Counting follicles takes magnified clinical images, not a phone held at arm's length, and even with those images the best published model was wrong about a quarter of the time. That is the honest ceiling on 'AI counts your hairs'. What a photograph of your whole head can give you is visible density, meaning how much scalp shows through, which is a different and coarser reading. It is still useful, because it is the reading that changes over months and the one you can compare against your own photograph from last year.

Four things move a reading more than most people expect, and all four are under your control.

Light is the biggest. Yellow indoor light warms colour and hides dullness; a bathroom downlight throws hard shadows into a part and makes density look worse than it is. Daylight near a window, with the light on your face rather than behind you, is the fairest condition.

Product is next. Oil, gel and cream all change how light bounces, which is the exact thing being read. Freshly washed and heat-styled hair is also misleading, since it hides the frizz and the oil pattern the reading is trying to see. Wet hair is worst of all: it looks darker, thinner and straighter than it is, and it is why people are startled by their own hair in the shower. And any filter, beauty mode or portrait blur is a problem, because those smooth exactly the edges that carry the damage information.

You should know this before you upload anything anywhere, and most tools make you dig for it. Here is ours, plainly.

The photographs you send to the Kurli AI Specialist go to Google Gemini, which is the model that writes the plan. We keep a copy so results can be checked later, which matters when you come back with a follow-up question and we need to see the same hair the plan was written for. The photographs are never sold and never published. If you want them removed, email hello@kurli.me and they are deleted.

That is the whole arrangement. Our view is that any hair tool asking for a photograph of your scalp should be able to state its version in four sentences, and that having to search for it is itself an answer.

Treat the output as a set of observations from a careful stranger who has seen a lot of hair and has never touched yours. That is close to what it is.

Accept the visible readings at face value, because those are the ones photographs are good at: curl pattern, visible density, the regrowth line, the state of your ends, where oil sits. Check the invisible ones yourself, since porosity and elasticity take two minutes and a glass of water. And discount anything that claims a diagnosis, no matter how specific it sounds.

The last habit is the most useful one: take the same three photographs again in three months, in the same light, and compare. A single reading is an opinion. Two readings three months apart is evidence, and it is the kind neither a mirror nor a memory can give you.