What “accuracy” means for an age guesser
Researchers measure age estimators with mean absolute error (MAE): the average number of years between the guess and the person’s real age, ignoring whether the guess was too high or too low.
An MAE of 5 doesn’t mean every guess lands within 5 years. It’s an average. Some guesses will be off by 1 year and others by 10 or more. For one photo of one person, you can’t know in advance which kind of guess you’ll get.
The numbers for the model in AgeBooth
AgeBooth runs the open-source face-api.js age and gender model in your browser. Its author published these test results:
| Age range | 0–3 | 4–8 | 9–18 | 19–28 | 29–40 | 41–60 | 60–80 | 80+ |
|---|---|---|---|---|---|---|---|---|
| Average error (years) | 1.52 | 3.06 | 4.82 | 4.99 | 5.43 | 4.94 | 6.17 | 9.91 |
Source: face-api.js README. Overall average error 4.54 years. Results differ by test set, from 3.20 years on one to 6.54 on another.
Two things stand out. Error grows with age: past 60 the model misses by about 6 years on average, and past 80 by almost 10. And these are lab results on curated face datasets. A casual selfie in a dim kitchen is harder.
How the best systems compare
The US National Institute of Standards and Technology (NIST) runs an ongoing evaluation of commercial age-estimation software. In its 2024 report, the average error on a common set of visa photos had dropped from 4.3 years in 2014 to 3.1 years.
Visa photos are close to ideal: a straight-on face, even lighting, a plain background. Even with photos like that, today’s top commercial systems still miss by about 3 years on average. NIST also reports that error varies between demographic groups, so some people get less reliable guesses than others.
Why your result changes from photo to photo
The model only sees pixels. Anything that changes the pixels can change the guess:
- Lighting. Light from above deepens shadows under the eyes and around the mouth, which reads as older. Soft light from a window in front of you smooths them out.
- Expression. In a 2017 study by Ganel and Goodale, people judged smiling faces as older than neutral ones, probably because of the lines a smile makes around the eyes. Models trained on human-labelled photos can pick up similar cues.
- Distance and angle. A phone held close distorts features, and a low angle adds shadow under the chin.
- Image quality. Blur, heavy compression and tiny faces leave less detail to read.
- Filters and beauty modes. Smoothing filters remove texture and usually push the guess younger.
- Glasses, beards, makeup and hats. They hide or change the features the model relies on.
AgeBooth reads the largest face in the photo and rounds the estimate to a whole number.
Why you shouldn’t use it to check someone’s age
With an average error near 5 years, a 16-year-old and a 21-year-old can easily get the same guess. That’s why AgeBooth’s disclaimer says not to use results for identity, age checks or eligibility decisions. Systems built for age assurance use far more controlled photos, liveness checks and safety margins, and even they are judged against strict test conditions.
How to get a steadier guess
- Face a window so daylight falls evenly on your face.
- Hold the phone at eye level, at arm’s length or further, or use the 2x lens.
- Keep a relaxed, neutral expression.
- Turn off beauty filters and take off sunglasses and hats.
- Try three photos and look at the middle result, not the lowest one.
For more on picking a photo, see the photo guide.
Questions
Why did two photos of me get different ages?
Lighting, expression, distance and image quality all change what the model sees. Differences of several years between photos of the same person are normal for this kind of tool.
Does a younger guess mean I’m healthier?
No. The estimate reflects how a face looks in one photo. It isn’t a health measure or a biological age test.
Does AgeBooth store my photo?
No. The photo is analyzed in your browser and isn’t uploaded to an AgeBooth server.