Talk to usTry the demo

Face Age Estimation: How Accuracy Is Measured and How to Read the Numbers

Evaluation and benchmarks

For: Product owner, safety lead or procurement reviewer who has been shown an accuracy figure for age estimation from a face

What the US national measurement laboratory's age estimation evaluation reports, which error measures matter for an age gate, and what the international age assurance framework standard adds.

A supplier tells you that its software looks at a face and estimates the age of the person, and that it is "accurate to within a couple of years". Before you let that sentence decide who gets into an age-restricted service, you need to know what was measured, on which photographs, with which error measure, and for whom the figure holds. This article reads the public evaluation of age estimation software published by the US National Institute of Standards and Technology, and the first part of the international standard on age assurance, and turns them into questions you can put to any supplier.

It is educational. NeuroVision does not offer age estimation from a face; the age check that is part of its verification flow works on the date of birth, and the last section says how.

Two different tasks that share one name

An age assurance system can be built in two ways, and the evaluation keeps them apart. An age estimation algorithm emits a number, an estimated age, which can then be compared with a threshold. An age verification algorithm in the narrow sense of the evaluation emits only a yes or no answer to the question whether the person is above an age threshold. The report adds that the yes or no can be produced from an estimate, or computed without any explicit estimate.

The same words are used by regulators in a wider sense. In the United Kingdom, section 230 of the Online Safety Act 2023 defines age estimation as any measure designed to estimate the age or age-range of users of a regulated service, and age verification as any measure designed to verify the exact age of users. A Spanish joint post on the European guidelines describes age estimation as algorithmic methods, such as facial analysis or behavioural profiling, that establish a likelihood that a user falls within an age range or is above or below an age threshold. So a face-based method sits in the estimation family, however a product page names it.

The evaluation also separates the technique from face recognition. It says that age estimation analyses one face to produce an estimate of age, while face recognition is concerned with who is in an image, and that the two are trained on different data: photographs with known ages for estimation, pairs of photographs with identity labels for recognition. An age estimate typically needs one photograph or a video, and the photo can be deleted afterwards, which is why privacy regulators look at the technique with interest. The summary of the report states that age estimation can operate statelessly, with no requirement for persistent storage of a photo or of biometric data derived from it.

Who measured, and what the measurement is

The National Institute of Standards and Technology, the US national measurement laboratory, runs the Face Analysis Technology Evaluation, abbreviated FATE. Its track on age estimation and verification published its first report in May 2024, as Internal Report 8525, the institute's first on the topic in a decade. The institute last evaluated such software in 2014, in a one-time effort on a single database of about 6 million visa photographs.

The 2024 report evaluates six algorithms that developers provided voluntarily in response to a call for submissions in September 2023. The test photographs number about 11.5 million and come from four databases, all from United States government sources. They are the visa collection used in 2014, a set of FBI mugshots, a set of webcam images obtained at border crossings, and a set of immigration application photographs of people born in more than 100 countries. All data was anonymised, and the research was reviewed to protect the rights and privacy of the photographed subjects.

Three limits of the design matter for anyone reading a supplier's claim.

  • The study is a closed box study: the researchers analysed the end performance of the algorithms, not their inner workings or how they arrived at their results.
  • It excludes performance measured in interactive sessions, in which a person can cooperatively present and re-present to a camera, and it does not measure accuracy effects related to disguises, cosmetics or other presentation attacks.
  • It does not address policy and does not recommend thresholds, since these differ across applications and jurisdictions; the institute makes no recommendation on whether the software is fit for particular use cases.

The report also describes the six algorithms as prototypes, so not necessarily available as mature integrable products. A good result in the report therefore says something about a research snapshot of late 2023, and nothing by itself about the version a supplier ships to you now. The institute says as much about timing: the results are a partial snapshot of the field as it stood in late 2023, and the program accepts new submissions on a rolling basis, with updates planned every four to six weeks.

The error measures, one by one

The report defines its measures precisely, and a supplier should be able to name which ones its figure is.

Mean absolute error. For each test photograph the absolute error is the difference between the estimated age and the true age, and the mean absolute error, abbreviated MAE, is the mean of these over all photographs. The report calls it the primary accuracy measure. The phrase "accurate to within a couple of years" is usually a statement of this kind.

Median absolute error. The median of the same errors, which is insensitive to outliers.

Standard deviation. It quantifies variation around the mean, and the report treats it as a secondary indicator, because an algorithm could have a low standard deviation and still systematically underestimate or overestimate age, which means a large MAE.

Accuracy at a tolerance. The proportion of test images for which the absolute error is below a chosen limit.

Failure to process. The share of photographs the algorithm could not process at all; the report lists it next to the error measures.

The report is blunt about the limit of the headline measure. MAE does not specifically quantify how many age estimates are large overestimates, and the report concludes that it is not an appropriate metric for age verification tasks. An error of a few years is harmless when the person is far above the threshold, and decisive when the person is just below it. That is the reason for a second family of measures built around a threshold.

The challenge age: how an estimate becomes a gate

A face-based gate is rarely used as a plain cut-off at the legal age. The report describes the usual configuration: because estimates contain errors, the system requires another form of age verification if the estimated age is below a challenge age, written T, that is set above the legal limit, written L. For a legal limit of 18, a challenge age of 25 is common, according to the report. For a legal limit of 21 a higher challenge age would be required to reach the same accuracy.

Two error rates follow, and both are about people, not photographs.

  • A false positive occurs when the true age is below the legal limit and the estimated age is at or above the challenge age: a minor who passes the gate.
  • A false negative occurs when someone above the legal age is estimated to be below the challenge age: an adult who is sent to another check.

The report builds two aggregate measures on these. Ineffectiveness is a population-weighted and risk-weighted average of the false positive rate, and it quantifies how well the system prevents people below the legal age from reaching the restricted item or service. Inconvenience is the matching weighted average of the false negative rate, and it quantifies how often people above the legal age are challenged and delayed. The report is open that its weights are illustrative: the likelihood that a person of a given age tries to defeat the system is modelled with an ad hoc exponential curve, under which 60% of 17-year-olds would try and only 1.1% of 13-year-olds, and it presents the model as something a local policy analyst should adapt or replace.

The consequence for a buyer is a trade-off that no supplier can remove. Raising the challenge age lowers the number of minors who pass, and raises the number of adults who must produce a document. The report's own tables show it. Across the six algorithms, the weighted inconvenience for a legal limit of 18 and a challenge age of 25 ranges from 0.082 to 0.122, and at a challenge age of 28 it ranges from 0.137 to 0.209. In the same setting the weighted ineffectiveness at a challenge age of 25 ranges from 0.012 to 0.220, and at a challenge age of 28 from 0.004 to 0.066. The spread between algorithms is as large as the effect of moving the threshold, which is why a figure without the algorithm, the threshold and the dataset cannot be compared with anything.

What the numbers looked like

The institute reports a gain in the headline measure since 2014. On the common database of visa photographs, the smallest mean absolute error fell from 4.3 years in 2014 to 3.1 years, and five of the six algorithms beat the algorithm with the smallest error that was submitted in 2014. In the summary table for subjects aged 18 to 24, the mean absolute error of the six algorithms still varies between algorithms on any one dataset and between datasets for any one algorithm, which the report attributes to sensitivity to image quality and to demographics.

The report's central finding is a negative one. No single algorithm stands out on every factor, and algorithm rankings differ across sex, image quality, region of birth, age itself and the interactions between those factors. Accuracy also depends on the metric chosen: an algorithm can lead on mean absolute error and not on the false positive rate at a challenge age.

A few specifics show how far the picture changes with the question asked.

  • For Application images, the false positive rate of 14 to 17 year olds at a challenge age of 25 ranges, across the six algorithms, from 0.006 to 0.248 for males and from 0.033 to 0.178 for females.
  • The same rate climbs steeply with the true age of the person below the limit: for one algorithm the false positive rate at a challenge age of 25 is 0.02 for 14 year olds and 0.295 for 20 year olds, almost fifteen times larger.
  • Read the two figures against the right limit. The report computes false positives against two legal limits, 18 and 21. A 14 year old is below both, so that figure is a false positive for an under-18 restriction too, while a 20 year old is below the limit of 21 only, so the 0.295 figure matters for an under-21 restriction and is not a rate for minors. For an under-18 gate, the comparable figures are those for ages 14 to 17.
  • For four of the six algorithms the false positive rate is markedly lower in men than in women, the opposite holds for one, and one is close to parity.
  • Four of the six algorithms show higher age estimation error, over or under, for both sexes when the person wears glasses than when not.
  • Children are easier in absolute terms: for the algorithm with the smallest error the mean absolute error is 0.34 years for children under 1, against 3.5 years for a 30 year old.

The report adds that error rates were almost always higher for female faces than for male faces, as they were for the algorithms evaluated in 2014, and that the underlying reasons are unknown. For demographic groups it uses six regions of birth as the grouping, and it explains that face processing algorithms should be expected to be sensitive to many features of a face, which is why it does not try to label images by appearance. To express how unevenly an error rate is spread over groups, the report borrows the Gini coefficient from economics and notes that the same approach has been used for biometric error rates in the 2024 edition of part 10 of the 19795 series of standards.

One more finding is easy to miss. The report's application programming interface includes a dedicated age verification function, and the report finds that the dedicated classifiers do not give a smaller error than the estimation algorithm compared with a threshold. A supplier who sells a separate "adult or minor" model as a gain in accuracy should be asked for the comparison.

Variation inside one person

The institute also shows a picture that numbers hide. For a staff member of the institute who changes facial expression, or puts on and takes off glasses, all six algorithms give age estimates that vary around the person's true age. With frames extracted from a cellphone video of a subject whose true age was 58, the estimates stay above or below the true age and vary by a few years from frame to frame. The report plots this for 1688 frames taken from a video of about a minute. For a gate that decides on a single frame, this frame-to-frame variation is part of the error; for a gate that decides on a stream, the supplier has to say how frames are combined.

What the international standard adds

The first part of the international age assurance standard, numbered 27566-1, was published by ISO and IEC in December 2025 as Edition 1, with 29 pages, by the joint technical committee JTC 1 subcommittee SC 27. Its abstract says that it establishes a framework for age assurance systems and describes their core characteristics, including privacy and security, for enabling age-related eligibility decisions. The public preview of the standard shows its structure, and the structure is a useful checklist even without the full text.

  • Its overview separates age verification methods, age estimation methods, age inference methods and successive validation.
  • A clause on functional characteristics covers sources of data, binding of the age assurance result to the correct individual, configuration management, context in use and delivery of the result.
  • A clause on performance characteristics covers effectiveness, indicators of effectiveness, performance metrics, resource utilisation and testability, and under metrics it lists classification accuracy, primary metrics, outcome error parity and performance efficiency.
  • A clause on privacy characteristics covers privacy by design and default, data minimisation, avoidance of adding to a digital footprint, user awareness and audit logs.

The institute's 2024 report described the standard as a multipart one: part 1 a framework including vocabulary, part 2 on architecture and interoperability, and part 3 on benchmarking and analysis. Check which parts are published when you read a supplier's claim of alignment with the standard, because a framework part sets out vocabulary and characteristics and does not by itself define a test that a product can pass.

Regulators and trial organisers have turned to the same questions. An Australian trial of age assurance technologies, commissioned by the federal communications department and conducted by an independent provider, published its final report on 31 August 2025 in parts that assess age verification, age estimation, age inference, successive validation, parental control and parental consent solutions. The department's page also lists a corrigendum to the trial issued in 2026, which is a reminder to check for corrections to any report you rely on.

Questions to put to a supplier

The sources above turn into a list that takes one meeting to run.

  1. Which error measure is the quoted figure: mean absolute error, median, accuracy at a tolerance, or a false positive rate at a challenge age?
  2. For a threshold use, what are the false positive rate and the false negative rate at your intended legal limit and challenge age, and what are the weights behind any aggregate?
  3. On which photographs: collected how, with which cameras, of which ages, and how close to the threshold? A test set with few people just under the limit says little about the cases that matter.
  4. Were the results broken down by age, sex and region of birth, and how large is the spread between groups?
  5. Is the figure for the version you will deploy, and who ran the test: the supplier, or an independent evaluator with a published protocol?
  6. How were poor-quality images, glasses and failures to process counted?
  7. What happens to a person who is challenged: which fallback check applies, and what does a person without a suitable document do?
  8. What does the service keep: is the photograph deleted after the estimate, and does the supplier keep anything derived from it?
  9. Is the figure measured on a cooperative live capture or on stored photographs, and is the system tested against someone who presents a photograph, a mask or a synthetic face?

The last two questions come from the gaps in the design of the evaluation rather than from the report's results: it measured stored photographs, not live sessions, and not attacks.

Where this sits in a verification flow

NeuroVision does not estimate age from a face. Where an age rule applies to a verification session, the face match step compares the selfie with the portrait on the document in a one-to-one comparison, and includes foreign-face detection and an age check against the date of birth. The date of birth is read from the document and compared with the age window of your scenario. That is a different technique with a different error profile, because the age comes from the document and not from the appearance of the person. For the regulatory frame around such checks, see the article on age assurance and age verification and the one on reading a face recognition evaluation.

To see the age check in your own scenario, contact the team.

Sources

All sources were fetched on 27 October 2026.

Ready to talk?

Tell us what you are building and we will help you plan the integration.

Talk to us