What the evidence says about age verification

A vendor says its facial age estimation has a margin of error of about 1.4-year [claim: roblox-age-estimation].

A regulator says the share of UK children encountering highly effective age checks rose from 25% to 43% between July 2025 and January 2026 [source: ofcom-age-assurance-2026].

A government-funded trial says a 12-year-old in its test would have been admitted to a 13+ tier in 62.89% of observations [source: ageassurance-aatt-age-estimation-2025].

A NIST benchmark says facial age estimation algorithms in its 2024 cohort had a mean absolute error ranging from 2.17 to 5.17 years at age 13 [source: nist-ir8525-2024].

All four statements can be true at once.

None of them, on its own, describes the probability that a specific child is on the correct side of a specific age boundary on a specific service.

This page is about what the evidence does and does not say about age assurance.

It is built on a research pass that located 13 independent evidence sources — government trials, regulator reports, NIST benchmarks, and peer-reviewed studies.

The companion page How age checks work — and how well covers the methods themselves.

The Roblox-specific material lives on Accounts, age limits and what a birthday changes.

Definition

A research-grounded synthesis of what independent evidence has measured about age assurance.

This is primarily facial age estimation, but also the broader regime of age checks, thresholds, and appeals.

It covers what the evidence has and has not measured.

It is not a vendor scorecard.

It does not name a “best” or “worst” provider.

It does not say whether age assurance works.

It says what the evidence base contains, what it does not contain, and where the disagreements lie.

Scope

This page covers the public, locatable evidence retrieved on 2026-08-19 on age assurance for users under 18 [source: ageassurance-aatt-age-estimation-2025].

The central document is the Australian Age Assurance Technology Trial (AATT) Part D.

It was published 2025-08.

It is the most direct independent test of age estimation at the legal thresholds 13+ and 16+ this wiki has located.

It also covers the NIST Face Analysis Technology Evaluation: Age Estimation and Verification (FATE AEV) rolling report.

It covers the UK Ofcom 2026 statutory review and the European Commission’s age-verification blueprint.

It covers the UK ICO’s age-assurance opinions and the Australian eSafety Commissioner’s regulatory guidance.

It also covers a small set of peer-reviewed research on algorithmic bias.

The page does not cover:

• Vendor marketing or product launches. Vendor claims appear here as [claim] and are not used as evidence of accuracy.
• The legal basis for age of digital consent. That is on the jurisdictional pages, not here.
• The accuracy of identity-document verification. The peer-reviewed evidence on document verification was not the subject of this research pass and is not in scope.

Why this matters

Parents of children between 6 and 16 are repeatedly exposed to two narratives about age verification.

The first is the marketing narrative.

A vendor’s age check “ensures safety” with a stated accuracy.

A single bad actor is treated as if it represents the whole category.

The Roblox public statement of “approximately 1.4-year” margin of error [claim: roblox-age-estimation] is the most prominent current example.

The second is the panic narrative.

Age checks are “fundamentally broken” or “surveillance”.

The published characterisation of facial age estimation in the AATT report — substantial boundary errors at exactly the ages that matter [source: ageassurance-aatt-age-estimation-2025] — is the most prominent current example.

Both narratives, taken at face value, overstate what the evidence can support.

The marketing narrative overstates by generalising from one system under one test protocol to the population of all children on all services.

The panic narrative overstates by conflating an average error with the probability of a rights boundary mistake for one child.

The honest reading is more specific than either:

• Independent cross-vendor benchmarks exist, and they are large [source: nist-fate-aev-2026].
• Independent threshold-specific trials exist, and they find substantial error immediately around 13 and 16 [source: ageassurance-aatt-age-estimation-2025].
• No independent evaluation located links a specific commercial system in production to a specific NIST submission or AATT result, for Roblox or any other named platform.
• Vendor accuracy figures are not falsified by the independent evidence, but they are also not validated by it.

The practical consequence for a parent is concrete.

A child who is admitted to a tier they should not occupy, or excluded from a tier they should, is the experience the evidence base has difficulty quantifying.

The right place to find that quantification is the production telemetry of the deploying service.

That telemetry is not in the public record this wiki retrieved for any of the major platforms.

What the research actually measured

The evidence base is a stack of four different kinds of study.

Each asks a different question.

Reading a number from one study as if it answered a question from another is the most common mistake parents make with age-assurance research.

Government-funded field trials with real users.

The Australian Age Assurance Technology Trial recruited children and adults.

It gathered ground-truth age from school rolls and consent processes.

It ran them through a panel of age assurance providers.

It published per-true-age FPR and FNR at the policy gates 13+, 16+, 18+ and 21+ [source: ageassurance-aatt-age-estimation-2025].

This is the strongest design located in the public record.

It has real users, real cohorts around the boundary, and gate-specific metrics.

The trade-off is that the trial tests participating providers under its protocol.

It does not test the same providers’ deployed production configurations in third-party services.

Cross-vendor algorithm benchmarks.

The NIST FATE AEV program compiles submitted algorithm libraries from commercial and academic developers.

It runs them on a large set of operational images.

The images are drawn from US arrest records, immigration applications, Mexican visa applications and US border-crossing photography [source: nist-fate-aev-2026].

The UK government describes the corpus as roughly 11.5 million images of more than 8.5 million individuals.

The current rolling report records 49 algorithm identifiers from 28 organisations, current to 2026-07-31.

This is the strongest cross-vendor benchmark located.

The trade-off is that NIST tests submitted libraries, not complete production age-assurance stacks [source: nist-ifpc-aev-2025].

Capture, liveness, retry, fallback, threshold policy and appeal flow are mostly outside the benchmark.

Regulator practice reviews.

The UK Ofcom 2026 age assurance review and the European Commission’s 2026 blueprint are not algorithm benchmarks [source: ofcom-age-assurance-2026] [source: digitalstrategy-ec-age-verification-2026].

They are analyses of how age assurance is being deployed, by whom, under what regulatory expectations, with what appeal and challenge design.

They report what services are doing, not what algorithms achieve.

Their value for the question on this page is the regulatory context and the small amount of service-level accuracy data they summarise.

Peer-reviewed academic studies.

A small number of peer-reviewed papers examine the methodology and bias of age estimation under controlled test conditions.

Hiba and Keller on the importance of subject-exclusive evaluation [source: ieeexplore-hiba-keller-2023].

Kärkkäinen and Joo on dataset bias and balanced Face Attribute data [source: openaccess-fairface-2021].

Clapes and colleagues on gender, ethnicity, makeup and expression as bias factors [source: openaccess-clapes-et-al-2018].

Han, Otto and Jain on human versus machine age estimation [source: ieeexplore-han-otto-jain-2013].

These studies contribute methodology and the demonstration that reported accuracy depends sharply on test protocol.

They do not evaluate specific commercial systems at the 13 or 16 boundary.

A claim in a vendor’s marketing deck is none of these.

It is not a government trial, not a NIST submission, not a regulator audit, and not a peer-reviewed study.

The peer-reviewed literature on the question of whether vendor self-reported accuracy matches the independent measurement at the relevant boundary is, in 2026, not yet substantial enough to draw a generalised conclusion.

The most direct threshold evidence: AATT

The Australian Age Assurance Technology Trial, commissioned and funded by the Australian Government, led by the Age Check Certification Scheme with KJR as testing partner, is the clearest public independent evidence located at the 13+ and 16+ boundaries [source: ageassurance-aatt-age-estimation-2025].

The trial’s pooled-sample FPR and FNR for the 13+ gate are the most quoted figures in any recent age-assurance debate.

The numbers this wiki cites are the ones in the published report’s main tables, with the sample sizes the report provides:

[source: ageassurance-aatt-age-estimation-2025]

[source: ageassurance-aatt-age-estimation-2025]

The trial’s overall best-system mean absolute error is reported at around 1.3 to 1.5 years [source: ageassurance-aatt-age-estimation-2025].

The gap between a 1.3-to-1.5-year overall MAE and a 62.89% false positive rate immediately below the 13+ gate is the core finding.

A small average error and a substantial boundary error are not the same measurement.

NIST is explicit that mean absolute error alone is not an adequate metric for age verification.

It can conceal the direction and distribution of errors at the boundary [source: nist-fate-aev-2026].

AATT also reports that its trial’s aggregate fairness analysis found broad demographic groups within its selected fairness criterion.

This finding, which is part of the same report, is more encouraging than the boundary error tables alone.

The two findings are not contradictory.

The trial’s own narrative says its aggregate fairness analysis should be read alongside, not as a replacement for, threshold-level analysis [source: ageassurance-aatt-age-estimation-2025].

Caveats specific to the AATT report.

The published report contains consequential internal inconsistencies the reader should know about:

• The 16+ numerical table reports FNR 13.90% for age 16 and 8.66% for age 17, while nearby prose gives 8.5% and 2.6% for the same ages [source: ageassurance-aatt-age-estimation-2025]. The table and the prose cannot both be right. The report also gives TPR 93.71% at age 19 in the table and describes TPR as consistently above 95% from age 19 in the prose.

• Two providers were excluded from the pooled 16+ sample “because of consistently poor performance” that the report said would unfairly skew the aggregate. The pooled result is therefore not an all-participant result.

• The report’s own summary of facial-image test counts varies between 27,457 in one place and 28,514 in another.

• AATT does not provide a complete skin-tone-specific FPR/FNR matrix at the 13+ and 16+ gates. The aggregate Fitzpatrick group MAE values are 2.55 years for types 1–2, 2.60 for types 3–4 and 3.01 for types 5–6 [source: ageassurance-aatt-age-estimation-2025]. These are population MAE values, not gate-conditioned FPR/FNR.

The wiki cites the table values as published, with the discrepancies named.

A corrected erratum, provider-level anonymised outputs or the raw trial dataset would resolve some of these uncertainties.

As of 2026-08-19, they are not public.

What AATT does not tell you.

AATT does not map its measured error rates to any specific deployed commercial system.

AATT does not measure whether a child who was misclassified in its test successfully appealed, abandoned the process, or remained in the wrong age band.

AATT does not test liveness, full image-quality robustness, or the real-world device/capture variation of a child’s phone in their bedroom.

Those are open questions, not answered questions.

The largest cross-vendor benchmark: NIST FATE AEV

NIST’s Face Analysis Technology Evaluation: Age Estimation and Verification program is the most rigorous large-scale independent comparison of age estimation algorithms available in the public record [source: nist-fate-aev-2026].

The current rolling report records 49 algorithm identifiers from 28 organisations, with submission dates ranging from 2023 to 2026.

NIST’s original 2024 evaluation on the Mexican visa dataset gives MAE values at specific ages, which is the most direct comparison this wiki has located:

• At age 13, MAE ranged from 2.17 to 5.17 years across the six algorithms initially evaluated [source: nist-ir8525-2024].
• At age 16, MAE ranged from 2.73 to 4.96 years across the same six algorithms [source: nist-ir8525-2024].
• The age-13 and age-16 samples in that evaluation contained 94,047 and 96,388 Mexican-visa images respectively [source: nist-ir8525-2024].

The NIST result that the headline is most likely to miss is a methodological one.

NIST states explicitly that mean absolute error alone is not an adequate metric for age verification [source: nist-fate-aev-2026].

It can conceal the direction and distribution of errors relevant to an eligibility boundary.

NIST’s current report also identifies 13–16-year-old online-safety applications as future work.

The rolling programme does not yet publish a 13+ or 16+ gate FPR/FNR equivalent to AATT’s threshold tables [source: nist-fate-aev-2026].

NIST’s analysis also finds that age-estimation accuracy depends on algorithm, sex, age, image quality and region of birth [source: nist-ir8525-2024].

The mean age-estimation error was generally higher for women than men across the evaluated datasets and age groups in its initial evaluations.

NIST uses region or country of birth as a demographic proxy.

It explicitly cautions that it is not a direct measure of ethnicity [source: nist-fate-aev-2026].

What NIST does not tell you.

NIST evaluates submitted algorithm libraries under a common test harness.

NIST therefore has more control over the tested executable than a survey-based regulator does.

NIST is still testing a submitted algorithm, not the complete production service with camera UI, liveness, retries, threshold policy and appeal flow [source: nist-ifpc-aev-2025].

A “NIST-tested” result is not a certification of the deployed system [source: nist-fate-aev-2026].

NIST does not certify that a product using the same company name has the same performance as the submitted build.

Results should be cited with algorithm ID and report date rather than summarised as “NIST-certified accurate”.

Demographic bias: what the evidence does and does not say

This is the part of the evidence base where the most caution is warranted.

NIST finds performance depends on algorithm, sex, age, image quality and region of birth [source: nist-fate-aev-2026].

Its demographic analyses do not support treating a single global MAE as evidence of demographic parity.

NIST treats region of birth as an imperfect proxy for ethnicity, not a direct one.

AATT reports aggregate MAE of 2.55, 2.60 and 3.01 years for Fitzpatrick skin-type groups 1–2, 3–4 and 5–6 respectively [source: ageassurance-aatt-age-estimation-2025].

The Fitzpatrick scale is a six-point skin-type scale, not a biometrically well-defined race or ethnicity category.

The AATT aggregate result is consistent with the trial’s narrative that demographic effects exist [source: ageassurance-aatt-age-estimation-2025].

The report itself does not provide the gate-specific FPR/FNR matrix by skin type that the question of demographic fairness at the 13 and 16 boundaries requires.

The peer-reviewed literature on dataset and model bias contributes the methodological point.

FairFace documented substantial demographic imbalance in commonly used face datasets.

It constructed a 108,501-image dataset designed to be more balanced across seven race groups [source: openaccess-fairface-2021].

Clapes and colleagues documented that gender, age, ethnicity, makeup and expression can affect age estimation [source: openaccess-clapes-et-al-2018].

Hiba and Keller showed that the choice of test protocol — random image split versus subject-exclusive split — can change a benchmark’s reported MAE from approximately 1.13 years to 2.53 years on the same dataset [source: ieeexplore-hiba-keller-2023].

Random splitting can leak identities between training and test sets.

The UK Information Commissioner’s Office’s 2025 equality assessment states, explicitly, that the ICO does not yet adequately understand whether protected groups are disproportionately affected by facial-age-estimation inaccuracies [source: ico-equality-impact-2025].

The Australian eSafety Commissioner’s regulatory guidance treats age assurance as not 100% effective, particularly close to a 16-year threshold [source: esafety-codes-guidance-2026] [source: esafety-smma-guidance-2025].

The UK Ofcom review found that only two of the services it examined reported technical accuracy metrics broken down by demographic characteristics [source: ofcom-age-assurance-2026].

Several depended on provider testing without service-level evidence sufficient to assess representativeness.

The honest reading: demographic effects are documented at the population MAE level for some systems.

Their magnitude at the 13+ and 16+ gate FPR/FNR level is not.

The regulator side, in the UK, in Australia, and at the ICO, has not claimed it knows.

The regulatory context

The UK Online Safety regime is the most concrete example of a jurisdiction treating age assurance as a regulatory question, not a marketing question.

Ofcom’s 2026 statutory review reports that across a sample of 32 UK services, more than 69 million age checks were completed between July and December 2025.

That is a 23-fold increase on the prior six months.

The same report finds that the share of children encountering “highly effective” age checks rose from 25% to 43% between July 2025 and January 2026 [source: ofcom-age-assurance-2026].

Ofcom’s review is not an algorithm benchmark.

It is a review of the regulatory practice of age assurance on deployed services.

It is the source of several service-level findings relevant to the question on this page:

• The most common challenge age used by facial-estimation providers in the examined adult-gate applications was 21, with a range of 20 to 25 [source: ofcom-age-assurance-2026]. Challenge ages are deliberately set above the legal boundary to manage false-accept risk.

• Appeal design across services is heterogeneous. Some services offered manual review or alternative methods. Some offered only another attempt using the same method. Some offered no appeal at all. Only six of the examined services supplied data on upheld appeals [source: ofcom-age-assurance-2026].

• Self-declaration alone is not treated as effective age assurance under the UK regime [source: ofcom-age-assurance-2026].

• Service-level demographic accuracy data is sparse enough that two services in the examined sample provided demographic breakdowns. Several relied on provider testing rather than service-level testing [source: ofcom-age-assurance-2026].

The European Commission’s 2026 age-verification blueprint is not a benchmark either.

It is a privacy-preserving architecture [source: digitalstrategy-ec-age-verification-2026].

A user proves to a relying service that they are over an age threshold, without disclosing the underlying identity or other attributes to the relying service.

The blueprint was first released on 2025-07-14 and reached a feature-ready version on 2026-04-15.

Its existence is the most important piece of context for the data-minimisation question.

It is not evidence about the accuracy of any particular age estimation system.

The UK Department for Science, Innovation and Technology’s 2026 data-access study is the most recent UK synthesis of the technical state of age assurance [source: www-uk-age-assurance-data-access-2026].

It identifies data access and representative training and testing data as material constraints on age-assurance performance.

It refers to NIST as the largest public benchmarking basis.

It notes the difficulty of narrow age bands such as 13–15 and 16–17.

The UK ICO’s age-assurance opinion requires organisations processing children’s data to consider fairness, accuracy, data minimisation, purpose limitation, retention and security [source: ico-biometric-age-assurance].

A Data Protection Impact Assessment is required for high-risk processing.

The ICO’s analysis treats the boundary between “ordinary” and “special category” biometric processing as purpose-dependent.

Facial age estimation is not automatically special-category under UK GDPR merely because a face is processed [source: ico-biometric-age-assurance].

What we do not know

Some of the most relevant questions for a parent are unanswered in the public evidence base as of 2026-08-19.

No independent validation of a specific production system.

No public source located in this research pass maps the facial-age-estimation implementation currently used by Roblox, or any other named deployed service, to a specific NIST algorithm identifier and version.

Provider-level NIST or AATT results cannot responsibly be attached to a specific deployed system until this mapping is established.

A vendor accuracy figure of “about 1.4-year” [claim: roblox-age-estimation] is therefore a manufacturer claim.

It is not validated by the located independent evidence.

No replicated boundary error rate for any named commercial system.

AATT supplies independent reference values at 13+ and 16+: FPR 62.89% for age 12, FNR 23.57% for age 13, FPR 73.26% for age 15, FNR 13.90% for age 16.

These are for AATT’s pooled test configuration.

A product-specific assessment should report confidence intervals and distinguish exact age in months.

A newly 13-year-old and a nearly 14-year-old need not have equal error probability.

No gate-specific FPR/FNR by sex, skin tone or intersectional group.

AATT’s aggregate Fitzpatrick MAE values are 2.55, 2.60 and 3.01 years for grouped skin types 1–2, 3–4 and 5–6 [source: ageassurance-aatt-age-estimation-2025].

The gate-specific FPR/FNR matrix by skin type at 13+ and 16+ is not published.

NIST’s sex-based analysis finds a generally higher MAE for women [source: nist-ir8525-2024].

It does not publish gate-specific 13/16 FPR/FNR by sex.

The UK ICO’s equality assessment states that the regulator does not yet adequately understand the question [source: ico-equality-impact-2025].

No independent data on appeal success, abandonment or correction time.

Ofcom’s evidence base for the UK shows that appeal design is heterogeneous and sparsely measured.

Only six of the examined services supplied upheld-appeal data [source: ofcom-age-assurance-2026].

The number of children who successfully correct a misclassification, the number who abandon the process, and the time taken are not in the public record this wiki retrieved for any major deployed platform.

No replicated test of image-capture robustness at production scale.

NIST’s analysis includes image quality as a determinant of accuracy.

AATT’s methodology intentionally included real-world capture variation [source: ageassurance-aatt-age-estimation-2025].

What is not in the public record is a product-specific test of camera resolution, illumination, pose, motion blur, glasses, occlusion, device generation and compression at the 13 and 16 gates for a specific deployed system.

No data on what the harms actually are.

A false positive admits an under-age user to a tier they should not occupy.

A false negative denies an eligible user.

The technical consequence is well-understood.

The safety consequence depends on which contacts, communication functions, content and discovery mechanisms that tier permits.

The peer-reviewed literature on the causal relationship between an age-misclassification event and measurable downstream harm was not located in this research pass.

What an independent third-party test would have to publish to close the gap.

A minimum publication package would have to include: exact production model and version.

An independent third-party test commissioned under terms allowing full publication.

A representative and independently sourced ground-truth cohort with oversampling around the boundary years [source: nist-fate-aev-2026] [source: ageassurance-aatt-age-estimation-2025] [source: ofcom-age-assurance-2026] [source: ico-equality-impact-2025].

Exact age in months.

FPR/FNR and confidence intervals at 9, 13, 16, 18 and 21.

Failure-to-acquire or failure-to-estimate rates.

Sex, pigmentation, ethnicity and intersectional analyses.

Camera, lighting and device robustness.

Liveness performance.

Challenge-age and retry logic.

Appeal outcomes.

Data-retention audit.

Repeat testing after material model updates.

How to read age-assurance claims

Five habits help a parent read past the headline.

1. Distinguish vendor claim from independent test.

A vendor accuracy figure is a manufacturer statement unless and until an independent test published it.

A “±1.4-year” or “99% accuracy” headline is the first thing to check, not the last.

2. Ask which metric the number is.

Mean absolute error and false-positive rate at a specific age gate are not interchangeable [source: nist-fate-aev-2026].

A 1.4-year MAE describes an average signed distance from a true age.

A 62.89% FPR at the 13+ gate describes the probability that a 12-year-old crosses the boundary.

NIST explicitly warns that MAE is poorly suited to describing age-verification decisions.

3. Notice the boundary that matters for the user.

A “1.5-year error” at the population level is consistent with much larger errors immediately around 13 or 16 [source: ageassurance-aatt-age-estimation-2025].

AATT’s tables show this directly: a best-system MAE of 1.3 to 1.5 years sits alongside a 62.89% FPR at 12 against a 13+ gate [source: ageassurance-aatt-age-estimation-2025].

4. Read the protocol.

Hiba and Keller showed that a random-image-split protocol and a subject-exclusive protocol on the same data can give MAE values of 1.13 and 2.53 years respectively [source: ieeexplore-hiba-keller-2023].

The protocol is the difference between the two numbers.

A vendor “accuracy” without a stated protocol is not a falsifiable claim.

5. Notice the population the test was run on.

A NIST result on operational US immigration and border-crossing imagery [source: nist-fate-aev-2026] is not the same population as a 12-year-old in a school bedroom with back-lighting and a parent’s old phone.

Cross-population transfer is itself an open research question.

What a parent can act on

The evidence base does not support a single verdict on age assurance.

It supports three practical moves a parent can make on the child’s behalf:

• Take the method question seriously. Self-declaration alone is not treated as effective age assurance in the UK regime [source: ofcom-age-assurance-2026]. Where a service asks for an ID or a face, the parent — not the child — is the relevant party for consent under eSafety’s Australian Social Media Minimum Age guidance [source: esafety-smma-guidance-2025].

• Plan for the false-negative case. Ofcom’s evidence shows that a non-trivial share of age checks fail and require retries or alternative methods [source: ofcom-age-assurance-2026]. The correction path on each service needs to be a known path, not a discovery at the moment of exclusion.

• Do not transfer a vendor’s headline number to a different service. AATT’s pooled-sample figures are reference values for the Australian trial, not for any specific deployed system. NIST’s MAE values are for the specific algorithm identifiers and versions NIST evaluated. Neither figure is a per-service accuracy number.

Sources

The page draws on the following sources. None is a vendor accuracy claim.

Every cited study, trial, regulator report or government study is either a government-funded independent trial, a national-laboratory benchmark, a regulator’s statutory report, or a peer-reviewed study.

Government-funded independent trial:

• ageassurance-aatt-age-estimation-2025 — Australian Age Assurance Technology Trial, Part D - Age Estimation. Pooled 13+ and 16+ threshold evidence, best-system MAE, Fitzpatrick-group MAE, table-versus-prose discrepancies, and provider-pooling rules [source: ageassurance-aatt-age-estimation-2025].
• ageassurance-aatt-literature-2025 — Australian Age Assurance Technology Trial, Part K - Glossary, Bibliography and Literature Review [source: ageassurance-aatt-literature-2025].
• ageassurance-aatt-method-event-2025 — Australian Age Assurance Technology Trial, Stakeholder Engagement Event / evaluation methodology material [source: ageassurance-aatt-method-event-2025].
• ageassurance-aatt-sab-2025 — Australian Age Assurance Technology Trial, Stakeholder Advisory Board Minutes [source: ageassurance-aatt-sab-2025].

National-laboratory benchmark:

• nist-fate-aev-2026 — NIST, Face Analysis Technology Evaluation: Age Estimation and Verification, rolling report. 49 algorithm submissions from 28 organisations, current update 2026-07-31, the largest independent cross-vendor benchmark located [source: nist-fate-aev-2026].
• nist-ir8525-2024 — NISTIR 8525, Face Analysis Technology Evaluation: Age Estimation and Verification. The 2024 six-algorithm age-13 and age-16 MAE values [source: nist-ir8525-2024].
• nist-ifpc-aev-2025 — NIST IFPC 2025 presentation on FATE Age Estimation and Verification [source: nist-ifpc-aev-2025].
• nist-aev-publication-2024 — NIST publication record for the FATE Age Estimation and Verification program [source: nist-aev-publication-2024].

Regulator and government report:

• ofcom-age-assurance-2026 — Ofcom, Use of Age Assurance Report 2026. The 25%-to-43% rise, the 69-million-checks figure across 32 UK services July–December 2025, the challenge-age and appeal evidence, the demographic data sparsity [source: ofcom-age-assurance-2026].
• ofcom-age-checks — Ofcom, Age checks to protect children online. The parent-facing explainer [source: ofcom-age-checks].
• ico-biometric-age-assurance — ICO, Expectations for age assurance and data protection compliance. The DPIA requirement and the biometric-purpose analysis [source: ico-biometric-age-assurance].
• ico-equality-impact-2025 — ICO, equality-impact material addressing facial age-estimation inaccuracies [source: ico-equality-impact-2025].
• digitalstrategy-ec-age-verification-2026 — European Commission, EU age verification. The privacy-preserving blueprint, first version 2025-07-14, feature-ready 2026-04-15 [source: digitalstrategy-ec-age-verification-2026].
• esafety-codes-guidance-2026 — eSafety Commissioner, Online Safety Codes and Standards Regulatory Guidance [source: esafety-codes-guidance-2026].
• esafety-smma-guidance-2025 — eSafety Commissioner, Social Media Minimum Age Regulatory Guidance [source: esafety-smma-guidance-2025].
• esafety-log132-2025 — eSafety Commissioner, document set LOG 132 [source: esafety-log132-2025].
• www-uk-age-assurance-data-access-2026 — UK Department for Science, Innovation and Technology, Age Assurance Data Access Study [source: www-uk-age-assurance-data-access-2026].
• euconsent-gdpr-art8 — GDPR Article 8 [source: euconsent-gdpr-art8].

Peer-reviewed research:

• ieeexplore-hiba-keller-2023 — Hiba and Keller, Hierarchical Attention-Based Age Estimation and Bias Analysis, IEEE TPAMI. The 1.13-vs-2.53-year random-vs-subject-exclusive comparison [source: ieeexplore-hiba-keller-2023].
• arxiv-hiba-keller-preprint — Hiba and Keller full-text preprint corresponding to the peer-reviewed work [source: arxiv-hiba-keller-preprint].
• openaccess-fairface-2021 — Kärkkäinen and Joo, FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age for Bias Measurement and Mitigation, WACV [source: openaccess-fairface-2021].
• openaccess-clapes-et-al-2018 — Clapes and colleagues, From Apparent to Real Age: Gender, Age, Ethnic, Makeup, and Expression Bias Analysis in Real Age Estimation, CVPR Workshops [source: openaccess-clapes-et-al-2018].
• ieeexplore-han-otto-jain-2013 — Han, Otto and Jain, Age estimation from face images: Human vs machine performance, ICB [source: ieeexplore-han-otto-jain-2013].

Vendor sources cited only as [claim]:

• roblox-age-estimation — Roblox, Facial Age Estimation for Safer Chat [claim: roblox-age-estimation]. The vendor’s own description of the global deployment, the six age brackets, and the approximately 1.4-year margin of error.
• roblox-age-checks-2026 — Roblox Help, Understanding Age Checks on Roblox [claim: roblox-age-checks-2026]. The vendor’s documentation of correction routes.
• roblox-estimate-age-2026 — Roblox Help, How does Roblox estimate my age? [claim: roblox-estimate-age-2026]. The vendor’s description of the estimation flow.