Beyond accuracy: What NIST’s latest age estimation results mean for age assurance

The latest results of the National Institute for Standards and Technology (NIST)’s Face Analysis Technology Evaluation Age Estimation & Verification (FATE AEV) are up and add results from Ayonix and ioNetworks Inc.

NIST’s Face Analysis Technology Evaluation (FATE) Age Estimation and Verification (AEV) provides an independent view of how well one of those approaches — facial age estimation — actually performs. NIST evaluates algorithms against around 11 million photographs from four operational sources, including immigration applications, visas, border crossings and U.S. arrest mugshots.

Biometric Update’s recently published 2026 Age Assurance & Digital Age Credentials Market Report examines this emerging ecosystem, including facial age estimation, identity verification, digital age credentials and privacy-enhancing approaches.

Accuracy has several layers

NIST evaluates performance using several complementary metrics. Acc(3) measures how often an estimated age falls within three years of actual age. Mean absolute error (MAE) measures the average size of the error, while mean error (ME) shows its direction — whether people tend to be estimated older or younger.

For age verification, NIST also evaluates Challenge-T policies, where someone estimated at or above T can be accepted while those below it undergo additional age assurance. NIST notes T=25 is common for a legal age of 18, while a higher T would be required for a legal age of 21 to achieve the same accuracy.

The evaluation also considers demographic performance, failure to process (FTP) and computational resources across different image datasets.

Overall accuracy masks age differences

Our analysis, giving every age equal weight, puts Regula-000 first overall on Acc(3). But the leader changes considerably by age.

ROC-002 performs particularly strongly at ages 16–20 and Yoti-004 at 23–29, while Innovatrics algorithms lead through much of the 30s to 50s. Onfido-000 leads at 69–77 and Idemia-002 at 80–87.

MAE reinforces the age effect. Among leading algorithms, error is around 1.25–1.55 years for ages 0–17 but increases toward 3–4 years or more among older subjects.

For age assurance, however, an error close to a legal age can matter considerably more than the same error later in life.

ME shows why the direction of an error matters

MAE alone can hide an important distinction.

Estimating a 17-year-old as 20 and estimating the same person as 14 are both three-year errors and therefore identical from an MAE perspective. For age assurance, their consequences can be very different.

ME exposes that direction. Our analysis shows a broad tendency for younger subjects to be estimated older and older subjects younger. ROC-002 and Yoti-004, despite being among the accuracy leaders, show pronounced overestimation among teenagers. In practical terms, this means algorithms tend to block less teenagers in the immediate vicinity of age restriction thresholds.

Demographics add another dimension. East African subjects tend to be estimated older across leading algorithms, while East European subjects are generally underestimated or closer to directionally neutral.

For age assurance, how an algorithm is wrong can therefore matter as much as how wrong it is.

Thresholds and demographics change the risk

The chosen challenge threshold can substantially alter outcomes.

For a 20-year-old, Regula-000’s false-positive rate falls from 12.0 percent at Challenge-T 25 to 3.1 percent at T=28. ROC-002 falls from 14.0 to 3.6 percent and Yoti-004 from 16.4 to 3.7 percent.

A higher threshold can reduce false acceptance but can also send more legitimate users into supplemental age checks, creating a trade-off between risk and friction.

Demographic differences can compound that challenge. At age 24 and T=25, average acceptance across algorithms on Application images ranges from approximately 51 percent for Eastern European females to 70 percent for Eastern African males. On Border images, the lowest and highest demographic averages range from approximately 47 to 80 percent.

The image source itself therefore matters alongside age and demographics.

Accuracy comes at different operational costs

Algorithms close on accuracy can also differ substantially in speed.

Cognitec-000 takes approximately 292 milliseconds per age estimate, compared with 823 milliseconds for Regula-000, 1,037 milliseconds for ROC-002, 1,094 milliseconds for Yoti-004 and 1,405 milliseconds for Daon-002.

FTP adds another consideration: an algorithm can be highly accurate when it produces an estimate but still create friction if customers frequently require retries or fallback verification.

What this means for the market

NIST’s results challenge the usefulness of a universal age-estimation leaderboard.

For buyers, selection should consider accuracy at the ages that matter, magnitude and direction of error, performance around the required threshold, demographic consistency, capture environment, failure rates and speed.

For vendors, differentiation may increasingly move beyond headline MAE or accuracy toward demonstrating strong performance around commercially important thresholds, across populations and capture environments, without creating excessive friction.

The question is shifting from “Which algorithm estimates age most accurately?” to “Which algorithm produces the right age-assurance outcome for a particular population, threshold and deployment?”

Article Topics

biometric age estimation  |  biometric testing  |  biometrics  |  Face Analysis Technology Evaluation (FATE)  |  IDEMIA  |  Innovatrics  |  NIST  |  Onfido  |  Regula  |  ROC  |  Yoti