An online space meant only for people aged 13 to 16 needs a door that lets them in and keeps younger children and adults out. NIST’s Child Online Safety table measures how well a facial age-estimation model would guard that door. For our submission, bynn-000, the model that powers Agemin, NIST measured a 72.53% true-positive rate for ages 13 to 16. On that metric it ranks #1.
Bynn ranks #1 for true-positive rate in NIST’s Child Online Safety evaluation for ages 13-16
How NIST’s age estimation evaluation tests a model that guards a space for people aged 13 to 16, what the 72.53% true-positive rate counts, what it does not show, and what it means for Agemin and for enterprise buyers.
- Bynn Intelligence
- 5 min read
NIST FATE Age Estimation & VerificationNIST FATE Age Estimation72.53%true-positive rateKeep outLet inKeep out8131622
This article explains what NIST measured, how the evaluation works, what the result does and does not show, and what it means for an organization choosing age assurance as part of its enterprise AI stack.
The result
Bynn ranks #1 for true-positive rate in NIST’s Child Online Safety evaluation for ages 13-16. The evaluation is NIST FATE Age Estimation & Verification, the ongoing program in which the U.S. National Institute of Standards and Technology tests age-estimation algorithms submitted by developers. Our submission, bynn-000, was submitted on 17 June 2026. The figures here are from NIST’s results table, last updated 29 September 2026.
| Evaluation | NIST FATE Age Estimation & Verification |
|---|---|
| Table | Challenge-25 & Child Online Safety (13-16) |
| Dataset | Visa images (good quality) |
| Metric | True-positive rate, ages 13-16 |
| Result | 72.53% |
| Rank on this metric | #1 of 58 submissions |
| Submission | bynn-000, submitted 17 June 2026 |
| Results table | Last updated 29 September 2026 |
What NIST measured
NIST describes the table as accuracy measures for two common applications: Challenge-25 age restriction, and access for children aged 13 to 16 to an online chatroom. The Child Online Safety part models the second. A model is shown a face and estimates an age; the question is whether people inside the 13 to 16 band are placed inside it, and whether people outside it are kept out.
The true-positive rate
NIST defines the true-positive rate for this table as the proportion of subjects aged 13 to 16 whose age is correctly estimated as 13 to 16. In NIST’s words, true-positive rates “quantify how often children within age 13-16 are estimated to be within that age range.” A 72.53% true-positive rate means that, on NIST’s Visa image set, the model placed 72.53% of the people aged 13 to 16 inside the band.
How it was measured
FATE Age Estimation & Verification is open to developers worldwide. A developer sends NIST its algorithm as software, not its own results, and NIST runs it on NIST’s images. The Child Online Safety measures use NIST’s Visa dataset, which NIST describes as good quality. Because every submission is measured on the same images in the same way, the table can be read across submissions, which is what makes a rank on it meaningful.
Where the result sits in the field
The Age 13-16 true-positive rate column lists 58 submissions. bynn-000 is first at 72.53%. The next submission in the column reached 69.28%, and the values fall from there to below 10%. The spread is wide, which is a useful reminder that age estimation inside a narrow teenage band is a hard problem and that the choice of model makes a material difference.
The other measures in the table
The same table reports mean absolute error for the 13 to 16 band, false-negative rates (people aged 13 to 16 estimated as 8 to 12, or as 17 to 22) and false-positive rates for younger children and for people aged 17 to 22. Our rank is for the true-positive rate. On other measures in the same table, including the false-positive rates, other submissions perform better. We encourage anyone evaluating age estimation to read the whole table, not one column of it.
What this result is not
NIST evaluation is not certification or endorsement. The rank applies to one metric, the true-positive rate, in one category, Child Online Safety for ages 13-16. It measures age estimation within that band; it does not measure a reduction in online harm, and it is not a claim of overall leadership in protecting children or of universal superiority in age verification.
What it means for an enterprise buyer
For a service that has to admit an age band and exclude everyone else, the true-positive rate is the measure of friction for the people you want to serve. Every person aged 13 to 16 who is estimated outside the band is turned away, or sent to a slower and more intrusive check. A higher true-positive rate means fewer legitimate users meet that obstacle.
The false-positive rate is the other side of the same decision: how many people outside the band are let in. The two move together. A model, or a threshold, that admits more of the band can also admit more people outside it. A sound evaluation therefore asks three questions of any vendor:
- What are the true-positive and false-positive rates at the operating point you will actually run, not only at the most flattering one?
- Were they measured independently, on images the vendor did not choose?
- What happens to a user the estimate cannot place with confidence: a retry, another method, or a person?
An independent measurement gives the services that rely on a model something more useful than a vendor’s own claim: a public method, a public description of the dataset and a public table to check. It does not remove the need to test the model in your own flow, on your own users’ capture conditions, with your own fallback.
What it means for Agemin
Agemin helps digital services assess age and build age-appropriate experiences. Its offering spans facial and email-based age estimation, identity verification, and age-related content tools. Agemin brings Bynn’s facial age-estimation research, evaluated through NIST’s ongoing testing program, into practical age assurance. Because the offering includes more than one method, a service can pair facial estimation with another check for the users it cannot place. Read more on the Agemin page.
Limits and how to read the table over time
- NIST updates the table as new submissions are evaluated, so ranks can change. The figures here are as of the table dated 29 September 2026.
- The Visa images are good-quality photographs. Live captures in a real product vary in lighting, pose and camera, and results in production can differ.
- One metric is reported here. The full table and the report card for bynn-000 carry every measure NIST publishes for it.
Read the source
NIST publishes the evaluation and its tables at pages.nist.gov, and the report card for our submission at face.nist.gov. Our other results, each with its type of evidence, source and date, are on the research page. If age assurance is one part of a wider problem, see how we work with organizations on enterprise AI.
