How old do I look?

Updated 15 October 2026How we measure

What your result means

The number is the age your face reads as in this photo: what a model trained on tens of thousands of labelled faces guesses from the 112 by 112 pixel square around your face. Researchers call it apparent age, to separate it from your real age and from any idea of biological age. It answers the question you typed, “how old do I look?”, for one picture.

It comes with an error, and the error is large. On our own test the estimate was off by 7.9 years on average, so the result card shows a band of ±8 years around the number. About half the readings in our test landed inside that band and about half further out. A reading of 34 means “this photo reads as somewhere around 26 to 42, most likely in the middle”, not “you look 34”.

Three things about the number are worth knowing before you read much into it:

• It leans toward the middle. Like most age models, this one pulls answers toward the middle of the range it learned from. In our test it read people over 60 as about 8 years younger on average, and it read some people in their 20s as older.
• It depends on the photo more than on you. The same face in hard side light and in soft front light can come out several years apart. That is the reason for the photo tips below.
• It is not a judgement. Looking older or younger than your age in one photo says little about you. Even people judging faces in person disagree by years.

How the age estimate is made

Everything below happens in your browser, in about a second after the models have loaded:

1. Your photo is opened on your device and turned upright using its orientation tag.
2. Your face is found. Google’s MediaPipe Face Landmarker places 478 points on your face, the same model our other tools use. If it finds no face, more than one face, a face that is too small, or a head turned or tilted more than 15°, it asks for another photo instead of guessing.
3. The face is cropped. We take the box around your jawline and eyebrows, add a margin of a tenth of its size on every side, and scale it to a 112 by 112 pixel square. That is the crop the age model was built for, and you see it next to your result as “what the model saw”.
4. The age model reads the crop. It is a small convolutional neural network with about 420,000 learned weights, 430 KB in all. It turns the square of pixels into one number: an age in years.
5. It reads the mirror image too. The crop is flipped left to right and read a second time, and the two answers are averaged. A front camera that mirrors your selfie then gets the same answer as one that doesn’t.

The age model runs in plain JavaScript that we wrote for it, with no extra runtime. We checked it against the same network built with TensorFlow.js, the library face-api.js runs on: the outputs agree to within a thousandth of a year. Our other tools measure the geometry of your face instead; the methodology page explains those measurements.

The model and what it learned from

The age model comes from face-api.js, an open-source face analysis library by Vincent Mühler, released under the MIT licence. Its age and gender network is built for browsers: a compact version of the Xception design, small enough to load in under a second on a phone. We use only its age output. The network also has a gender output, which this page never runs.

According to its author, the network learned from eight public face collections: UTKFace, FG-NET, ChaLearn, the Wikipedia and IMDb portrait sets, CACD, MegaAge and MegaAge-Asian, each split 80% for training and 20% for testing. The noisiest of them were cleaned automatically first. These collections label age in different ways:

• Computed from dates. The IMDb and Wikipedia portraits (the IMDB-WIKI set, about half a million photos) take each person’s birth date and the year the photo was taken. Its authors say plainly that they cannot vouch for the labels: film stills can come from long productions, and dates can be wrong.
• Labelled by hand. UTKFace’s more than 20,000 faces, aged 0 to 116, carry age, gender and ethnicity labels that were estimated by an algorithm and checked by a person.
• Apparent age. The ChaLearn photos are labelled with the average guess of many people, not with the real age. That is closer to the question this page answers.

The author reports these mean errors, in years, on the held-out 20% of each collection:

That averages 4.5 years. But held-out photos from the same collections look a lot like the training photos: the same kinds of cameras, the same kinds of people, often the same people. Photos from outside those collections are harder, which is why we ran our own test.

How accurate it is: our test

On 15 October 2026 we tested the tool exactly as you use it, in a browser, on 22 photos of adults whose age on the day of the photo is documented: the birth date is public, and the photo’s own record gives the date it was taken. Half are official portraits that are in the public domain; half are openly licensed photos taken at public events. The set has 11 women and 11 men aged 23 to 84, of several ancestries and skin tones. Nobody in it is under 20. The full list, with each photo’s licence, is kept with our test notes.

The tool asked for a different photo for 6 of them, because the head was turned or tilted more than 15°. On the 16 it accepted:

• the mean error was 7.9 years, and the median error 8.4 years
• 7 of the 16 estimates were within 8 years of the real age
• across the whole set the estimates did not lean older or younger (on average they were half a year low), but they leaned strongly by age, as the table below shows

With the head-angle check switched off, all 22 photos give a mean error of 8.2 years. Broken down by age:

Across all 22 photos, studio portraits did better than event photos: a mean error of 7.2 years against 9.2. Smiles, glasses and event lighting were common among the largest misses. A laughing 73-year-old man was read as 53, a 54-year-old woman photographed at an event as 35, a smiling 84-year-old wearing glasses as 70, and a smiling 35-year-old woman as 23. One studio portrait, a woman of 52, went the other way and was read as 65.

To put the number in context: the best commercial systems in NIST’s 2024 evaluation reached a mean error of about 3.1 years on visa photos, which are taken under strict, even conditions. A 430 KB model that runs on your phone is not in that class. And 16 photos is a small test: with this many, the true average error for photos like these could plausibly sit anywhere from about 5 to 11 years.

Where age estimates go wrong, and for whom

Every age estimator, human or machine, has blind spots. The published research is consistent on a few of them, and our test showed the same patterns where it was large enough to show anything.

• Age itself. Errors grow with age. Older faces vary more, and face collections hold fewer of them, so models pull older faces toward the middle and read them younger. People do the same: in the study by Voelkle and colleagues, the age of older faces was harder to judge than the age of younger ones.
• Sex. In NIST’s 2024 evaluation, error rates were almost always higher for women’s faces than for men’s, as they were in NIST’s 2014 evaluation, and the reasons are not known. Puc, Štruc and Grm found the same direction in two openly available age models: accuracy was consistently higher for men.
• Ethnicity and skin tone. The large public face collections are dominated by lighter-skinned faces. When Kärkkäinen and Joo built FairFace, a collection balanced across seven groups, a model trained on it was more accurate on new photos and more consistent across groups than models trained on the usual collections. For age in particular the evidence is mixed: Puc and colleagues found no consistent effect of race across their test sets, which suggests photo conditions matter as much as the label. NIST also lists region of birth among the factors that change accuracy.
• Expression and makeup. People most often underestimate the age of smiling faces (Voelkle and colleagues). Clapés and colleagues labelled 8,000 faces with their real age, the age people guessed, and their gender, ethnicity, makeup and expression, and found that some of these attributes consistently skewed the guesses; correcting for them made a model’s estimates more accurate. Smiles were common among our largest misses.
• Glasses and image quality. NIST lists image quality among the main factors and showed one person’s estimate moving in every algorithm it tested when they put on or took off glasses.

Our test has 22 photos, too few to measure any of these gaps on our own model with confidence, so we don’t claim a number for them. If the reading seems far off for you, the reasons above are the likely ones, and the photo tips below reduce the ones you control.

How to take a photo for a fair reading

The model only sees a small square of pixels, so everything that changes those pixels changes the answer. For the fairest reading:

• Soft, even light from the front. Face a window in daylight. Light from above or from one side deepens every line and shadow, which reads as older.
• Camera at eye level, at arm’s length or further. A phone held close enlarges the nose and narrows the face. A camera below your chin shows the underside of the jaw.
• Look straight at the camera. The tool asks for another photo if your head is turned or tilted more than 15°.
• A relaxed, neutral face. Smiling and laughing were behind our largest misses.
• Glasses off, hair off your face. Frames and fringes cover the area around the eyes and the forehead.
• No filter, no beauty mode. Most phone front cameras smooth skin by default, and a smoothed face reads younger. Turn it off, or use the rear camera.
• A sharp original. Screenshots and photos that have been sent through messaging apps are compressed, which blurs fine texture.

Then take two or three photos and compare. If they agree within a few years, that is your face’s reading; if they jump around, the photos are the cause.

Why some people look older or younger than their age

What makes a face read as older is well studied. In a study of Danish twins and British women aged 45 to 81, Gunn and colleagues found that wrinkles, grey hair, thinner lips and visible sun damage each went with looking older for one’s age. Genes and environment contributed about equally, and much of the variation stayed unexplained.

Looking old for your age can also mean something. In a study of 1,826 Danish twins aged 70 and over, Christensen and colleagues found that how old people looked in photographs predicted survival, even after accounting for their real age: within a pair, the twin who looked older was more likely to die first. That finding is about averages in older adults, judged by groups of nurses and lay assessors. It does not make a single photo-based estimate a health test, and this page is not medical advice.

Some of what changes with age is geometry: the jawline softens and the eyelids can become hooded. Our guides on hooded eyes and the jawline and gonial angle explain those changes, and mewing before and after looks at what the evidence says about changing the jawline.

What this tool is not

• Not an age check. An average error of several years is far too large to tell whether someone is over 18, and the tool was not built or tested for that. Don’t use it on other people.
• Not for photos of children. You need to be 18 or older to use it, and our test only included adults.
• Not a health or skin test. It says how old a face looks in one photo, nothing more.

More from the same photo

Your age reading is about how your face looks. Our other tools measure its shape: the attractiveness test scores six measured features, the face symmetry test compares the two sides of your face, and the true mirror shows your face the way other people see it. They all run on your device, like this one.