The query arrives with a selfie and an expectation: give me the number. A page returns an age, sometimes with a range, sometimes with advice attached. Run the same photo through another page and the number can move by years.
This page is not an age test. It explains what an estimator reads from a face, why the same face gets different answers, how wide the error runs in published evaluations, and which parts of a photograph change the reading.
What age estimators actually read
An age estimator does not measure time. It reads a face and compares the visible pattern with patterns learned from labelled faces. The cues named by the pages read for this page are surface cues: skin texture and tone, lines and wrinkles, soft-tissue volume around the cheeks and eyes, and the geometry of features such as the eyes, nose, mouth and jaw. One page lists wrinkles and fine lines, another lists skin, eyes, oval and hair, and another lists facial geometry and texture patterns.
None of those cues is a clock. Two people of the same age can differ by a decade in how old their skin looks, because sun exposure, smoking, sleep, weight change and genetics all leave marks that a photograph records. A single image is one moment under one light. The estimator turns that moment into a number, but the number describes the appearance in the image, not the time since birth.
It is not bone age either. Clinical bone age is read from a radiograph of the left hand and wrist, where the growth plates are visible, and scored against an atlas such as Greulich-Pyle or a maturity scale such as Tanner-Whitehouse. The method is used in children and adolescents, whose growth plates are still open; the Greulich-Pyle atlas itself was built from radiographs of children in Cleveland taken between 1931 and 1942, and the Tanner-Whitehouse 3 update was published in 2001. A face photograph cannot recover that number, even in principle, because the growth plates are not in the frame. Facial age estimation and bone age assessment are different measurements with different inputs and different clinical uses.
The six pages read for this page were the age-estimation pages among the top organic results in October 2026. Their descriptions differ in scope. One offers a photo upload or a 60-second live camera scan and also reports pulse-derived heart rate and HRV. One runs its model in the browser and says the photo never leaves the device. One returns a single best guess plus a likely range. One adds skincare and styling suggestions. One describes its output as apparent age, not birth date. One calls the result an estimate of perceived facial age and states that it is not a medical device.
Only one of the six gives a range alongside the single number. None publishes a model name, a training-set size or an error rate. Their photo-handling statements also differ: two say the upload is deleted after processing, one says the photo is never stored, one says it never leaves the browser, and two show no retention or deletion statement on the page read. The table records what each page says, not what a separate policy might say.
Six age-estimation pages read on 23 October 2026, with the date each page itself displays. The table records what each page says it reads, how it says the photo is handled, and what it returns. None of the six publishes a model name, a training-set size or an error rate.
| Page and date shown | What it says it reads | Photo handling as stated | What it returns | What it does not disclose |
|---|---|---|---|---|
| Longevity Cities; page metadata 7 June 2026 | Facial features from a photo or a 60-second live camera scan; the scan also reads pulse for heart rate and HRV | Photo is never stored; emailed results delete after 7 days | An estimated facial age and the gap to chronological age; the live scan adds heart rate and HRV | No model name, training data or error rate; calls itself not a medical device |
| How Old Do You Look; copyright 2026 | Eyes, nose, mouth and skin texture | Uploads deleted after processing; not used to train the AI; a shared result image kept for up to 30 days | An apparent age | No model name, training data or error rate |
| FaceAge.ai; no date shown | Skin tone, texture and facial shape | Processed instantly, never stored or shared | An estimated age; a paid report adds 19 measurements | No model name, training data or error rate |
| Homiwork; page metadata 27 May 2026 | Skin, eyes, oval and hair, zone by zone | No retention or deletion statement visible on the page read | Perceived age plus skincare and styling tips | No model name, training data or error rate |
| Face Shape Detector; page metadata 10 May 2026, updated 22 October 2026 | Facial geometry and texture patterns; the page says the neural network runs in the browser | Photos are never uploaded; nothing stored or transmitted | A single best-guess age plus a likely range | No model name, training-set size or error rate |
| BeautyPlus; copyright 2026 Pixocial Technology | Wrinkles, fine lines, skin tone, eyes, smile, jawline and skin texture | Calls itself privacy safe; no retention or deletion statement visible on the page read | Age estimates for one face or several faces in a group photo, plus skincare and styling tips | No model name, training data or error rate |
The six pages are the age-estimation pages among the top organic results for the query, read on 23 October 2026. Two app-store listings were left out because they do not describe a method. A photo-editor page was left out because the page read returned navigation copy and no age-specific method statement. Dates are the ones each page displays: three give a full date in page metadata, two give a copyright year, and one shows no date.
No account · No ad pixels · No photo upload
Detection runs on-device with MediaPipe. Nothing is uploaded — no account, no email, no storage.
Why the same face gets different ages
The query also arrives as "how old does she look", aimed at someone else in a photo. The mechanics are the same. An estimator returns its reading of one image, and there is no birth certificate in the frame to check it against. When two pages disagree, both numbers can be consistent with their own models and both can be wrong about the person.
The first source of disagreement is the model. Ganel and colleagues compared ten open-source age-estimation systems on the same set of faces. The average unbiased error ranged from 2.99 years for one model to 8.08 years for another, and the authors note that training data overlapped considerably across the set, so architecture and training objective accounted for much of the difference. A page that returns one number and does not name its model gives the reader no way to tell which end of that range produced it.
The second source is a systematic bias by age group. The same 2026 study documents what its authors call the Chronological Age Bias effect: faces of young adults are perceived, on average, about four years older than their chronological age, while faces of old adults are perceived about five years younger. The reversal appears in human observers and in AI systems, but the AI bias is larger. A 25-year-old and a 65-year-old can therefore both be pushed toward the middle, and a model can look reasonable on average while being badly off at both ends.
The third source is the observer. Clifford, Watson and White showed 3,948 passport images to 84 perceivers and found an average error of about eight years. Two biases explained almost all of the variance: an assimilative serial dependency, in which an estimate is pulled toward the age of the face seen immediately before it, and the same young-old reversal. Around 95% of the variance in their age estimates was captured by those two terms. Perceived age is not a fixed property that a viewer reads off a face; it depends on what the viewer saw a moment earlier.
The fourth source is the photograph and its preparation. A crop that includes more forehead, a tilted head, low contrast, a webcam instead of a studio camera, or a pair of eyeglasses can all shift the reading. NIST's 2024 evaluation found that eyeglasses changed the error direction or size for four of its six algorithms, and that border-crossing photos taken with a webcam were the hardest of its four image types for four of the six. The same face photographed twice under different conditions is not the same input.
The fifth source is the output itself. A single number implies a precision that the underlying model does not have. A range is more honest, but it is still the range of one model on one image. Two pages can both be internally consistent, both return a plausible age, and still disagree by a decade on the same photo.
Error ranges reported in studies
The largest public evaluation of age estimation is NIST IR 8525, published in May 2024. It reports six commercial algorithms applied to around eleven million photos drawn from four operational repositories: immigration visas, arrest mugshots, border crossings, and immigration office photos. On the 18-to-24 subset, mean absolute error ranged from 2.5 to 5.1 years across algorithms and datasets. On a common visa set, the best mean absolute error had fallen from 4.3 years in 2014 to 3.1 years. For the most accurate algorithm in the child set, error under age one was 0.34 years, against 3.5 years at age 30.
Those are group averages on specific datasets, not a promise about one photograph. The same report says accuracy is strongly influenced by algorithm, sex, image quality, region of birth, age itself, and interactions between those factors. It also says there is no uniformly superior algorithm, and that rankings differ across those factors. An algorithm that leads on average absolute error can lose on a threshold-based measure, because the two metrics answer different questions. NIST evaluates both age estimation and age verification, and a small average error does not guarantee a small false-positive rate at a legal age cutoff.
The demographic finding needs care. NIST reports results by region of birth as a proxy for ethnicity, and it states plainly that it does not attempt to recover skin tone from images, because that is technically problematic and because face algorithms may be sensitive to many phenotypes at once. It also explains why it does not use the Fitzpatrick skin type: the scale was designed for photosensitivity, is not an accurate measure of pigmentation, and was originally meant to be assessed by a clinician. The honest summary is therefore narrower than a slogan: published evaluations show variation by region of birth and image quality, but they do not support a simple claim that a given skin tone adds a fixed number of years of error.
Two other studies are sometimes used to fill that gap, and neither measures what the slogan implies. Hsieh and colleagues manipulated the colour content of the same faces and found that higher red-green contrast made a face look younger to human observers; they did not test commercial age estimators or skin tones. Panić and colleagues tested dataset composition in training: oversampling minority groups alone did not equalise error across ethnic groups, while reducing samples from overrepresented groups lowered the spread of mean absolute error. That is a result about training data, not about the accuracy of a deployed page.
Human error is not a clean benchmark either. Clifford's average of about eight years is a human baseline, not evidence that people can judge age accurately. Ganel and colleagues argue that the conventional mean absolute error is confounded by response bias: a model that systematically overestimates every face can produce a larger MAE without being less precise around its own perceived-age mean. Their unbiased measure, MAE-U, removes that confound. After the correction, human error in their experiments rose with age up to mid-adulthood and then fell, and humans were more accurate than AI for the oldest faces. In a 2022 comparison, AI was less accurate and more biased than human observers overall, and lost more accuracy on older faces, smiling faces and female faces.
The practical reading is that an error figure is a distribution, not a margin of safety. A model with a three-year average error can still be ten years off on a particular face. A page that prints no error figure at all has not told the reader whether its model sits near the good end of the published range or the bad end.
Lighting, expression, makeup
Image quality is the easiest effect to see and the easiest to control. NIST found that border-crossing photos taken with a webcam, with more variation in head orientation and lower contrast, were harder than office-collected application photos of the same people for four of six algorithms. Eyeglasses raised error for four of six. Several of the pages read here ask for a clear, front-facing photo in even daylight and warn against filters, sunglasses and heavy makeup. That advice is about making the input match the training distribution, not about changing the person.
Expression changes the reading. Ganel and Goodale found that smiling faces were perceived as older than neutral faces of the same people in the 20-to-39 group, an effect attributed to smile-related wrinkles around the eyes. The effect disappeared in the 60-to-79 group, where age-related wrinkles are already present, and in the 40-to-59 group it appeared only for male faces. A smile is not a neutral variable in an age estimate; it is a cue that the model or the viewer may weigh differently depending on the face.
Makeup changes it too, and the direction depends on age. Russell and colleagues showed photographs of women aged 20, 30, 40 and 50 with and without makeup. Makeup made the 40-year-olds, and especially the 50-year-olds, look younger. It made no measurable difference at 30, and it made the 20-year-olds look older. The authors suggest two routes: makeup alters visual cues such as skin homogeneity, facial contrast and feature size, and it also activates an association between makeup use and adulthood, which can push a young face upward. The result is a set of findings about perception, not a instruction about what anyone should wear.
Colour changes it as well. Hsieh and colleagues showed the same faces with red-green contrast increased or decreased and found that the increased-contrast version was judged younger. They also found that colour images produced less disagreement between observers than grey-scale images, which suggests that colour carries useful age information rather than mere noise. None of these experiments says that a filter, a smile or a makeup choice changes a person's age. They say that the photograph is the input, and the input is not neutral.
For anyone comparing pages, the implication is simple: compare like with like. Use the same photo, the same crop and the same lighting across pages, and expect a different photo to produce a different number. A page that asks for an unfiltered, front-facing image is trying to remove one source of variation. It cannot remove the model, the training data or the base rate of the face.
A dated comparison of studies
The table orders eight sources from August 2018 to May 2026 by publication date. Four are studies of human perception, one evaluates commercial algorithms, one is a training-data experiment, and two compare AI systems with human observers. The mix is deliberate: no single paper answers the query, and the disagreement between the tasks is part of the evidence.
The human studies establish that age reading is biased and context-dependent even before a model is involved. The NIST report establishes what a large, controlled evaluation of commercial systems can measure, and how much the aggregate hides. The training study shows that group differences can be moved by changing the composition of the data. The AI-versus-human studies show that a model can be more precise on average and still carry larger biases at the ends of the age range.
What the table cannot show is a ranking of the six consumer pages in the first table. None of them publishes a validation set, a model name or an error rate, and a count of pages would not create one. NIST is the closest thing to an audit in this set, but it evaluates submitted algorithms, not the pages that a person reaches from a search result. The useful comparison is therefore between kinds of evidence, not between brands.
Eight published sources, ascending by publication date. The third column names the task; the fourth records what the source measured; the fifth records the reported error or effect; the sixth records what the source does not establish.
| Date | Source | Task | What it measured | Reported error or effect | What it does not establish |
|---|---|---|---|---|---|
| 3 August 2018 (online) | Russell et al., British Journal of Psychology | Human perception; makeup | Age estimates for photographed women aged 20, 30, 40 and 50, with and without makeup | Makeup made 40- and 50-year-olds look younger, made no difference at 30, and made 20-year-olds look older | Not a test of any estimator, and not a rule about makeup use |
| 17 October 2018 | Clifford, Watson & White, Royal Society Open Science | Human perception | 3,948 passport images shown to 84 perceivers | Average error about eight years; serial dependence and the young-old reversal together explained around 95% of variance | A human baseline, not an accuracy claim for AI |
| 26 November 2021 | Ganel & Goodale, Scientific Reports | Human perception; smiling | Age estimates for smiling and neutral faces from age 20 to 79 | Smiling faces looked older at 20–39; the effect disappeared at 60–79 and appeared only for men at 40–59 | Does not measure model error; expressions were posed |
| 29 December 2022 | Ganel et al., Scientific Reports | AI versus human | A large sample of prominent age-estimation systems compared with human observers | AI was less accurate and more biased; it overestimated smiling faces more and lost more accuracy on older and female faces | Does not name a consumer page or estimate its error |
| 14 August 2023 | Hsieh et al., Scientific Reports | Human perception; colour | The same faces shown with red-green contrast increased or decreased | Higher red-green contrast made the same face look younger; colour reduced disagreement between observers | A perceptual effect, not a statement about skin tone or model bias |
| May 2024 | NIST IR 8525 | Commercial algorithms | Six algorithms, around eleven million photos from four repositories | MAE 2.5–5.1 years on the 18–24 subset; best 3.1 years on a common visa set; 0.34 years under age one versus 3.5 at age 30 | Not an audit of consumer pages; no single algorithm leads across groups |
| 28 July 2024 | Panić et al., Mathematics | Dataset composition | VGG19-type models on UTKFace (23,705 images) and APPA-REAL (7,691) | Oversampling minority groups alone did not equalise error; reducing overrepresented samples lowered the spread | A training experiment, not a deployed estimator or a skin-tone measurement |
| 11 May 2026 | Ganel et al., Scientific Reports | AI versus human | 240 faces aged 20–79, ten open-source models, human observers in two experiments | Conventional MAE is confounded by bias; on the unbiased measure models ranged from 2.99 to 8.08 years; humans were more accurate for the oldest faces | Not a ranking of commercial pages; the face set was Caucasian and mostly neutral |
Dates are publication dates. Russell et al. appeared online on 3 August 2018 and in the 2019 volume (110(1):87–100). NIST IR 8525 gives only May 2024. The table records what each source reports, not a ranking of consumer pages or a guarantee for any individual face.
A number from an age estimator describes a photograph, not a body. It comes from a model trained on other faces, and it moves with the model, the image and the group being photographed. Published error is an average across faces, not a limit for one face, and the six pages read here publish neither a model name nor an error rate.
If the number matters, read the retention line before the result and treat the output as one guess among several. This site does not provide an age estimator or an age number; its face-shape pages describe outlines and proportions, which are different measurements from the age a face appears to be.
Questions this page answers
What does an age estimator actually read from a photo?
It reads visible cues: skin texture and tone, lines and wrinkles, soft-tissue volume, facial shape, and the geometry of features such as the eyes, nose and jaw. The pages read for this page describe those cues in different words. None reads bone or growth plates, which a photograph does not show. The output is an estimate of apparent age, not a measurement of chronological age.
Why does the same face get different ages on different pages?
Because each page uses a different model, training set, crop and calibration. The 2026 comparison of ten open-source models found average unbiased errors from 2.99 to 8.08 years, and the models shared much of their training data, so architecture and training objective alone moved the result. A page that returns one number and does not name its model gives no way to tell which of those systems produced it.
How large is the error in published age estimates?
NIST's 2024 evaluation of six commercial algorithms reports mean absolute errors of 2.5 to 5.1 years on faces aged 18 to 24, with a best result of 3.1 years on a common visa set. Human viewers averaged about eight years in a 2018 study. Those are averages across faces, not limits for one photo: an individual estimate can be much further off, and the oldest faces are where some models lose the most accuracy.
Can a photograph measure bone age?
No. Clinical bone age is read from a radiograph of the left hand and wrist and scored against an atlas or a maturity scale; it is used in children and adolescents, whose growth plates are still open. A face photograph shows skin, soft tissue and the shape of features, not the state of a growth plate. Bone age and facial age are different measurements with different inputs and different clinical uses.
Does makeup or a smile change the age a system returns?
Published studies say yes for human viewers, and the direction depends on age. Makeup made 40- and 50-year-old women look younger, made 20-year-olds look older, and made no difference at 30. Smiling made younger faces look older, but the effect disappeared in the oldest group. NIST also found that eyeglasses raised error for four of six algorithms, and that border-crossing webcam photos were harder than standard portraits.
What happens to a photo uploaded to an age page?
It varies. One page says the photo is never stored and emailed results delete after seven days. One says uploads are deleted after processing and not used for training, though a shared result image is kept for up to 30 days. One says the photo never leaves the browser. Others say only that processing is private. None of the pages read here publishes a model name or an error rate, so the retention line is often the most specific statement on the page.