Is Age-Estimation AI High-Risk Under the EU AI Act?
Facial age estimation for age-gating is usually not Annex III high-risk — it is Article 50(3) biometric processing. The line, and when it turns prohibited.
Probably not. Facial age estimation — software that looks at a selfie and outputs an estimated age or age band to gate access to age-restricted content or products — is generally not a named high-risk use under Annex III. It is biometric processing that carries an Article 50(3) transparency notice and triggers GDPR Article 9 special-category rules, but it does not, on its own, land in the high-risk tier the way biometric identification or sensitive-attribute categorisation does.
Two things flip that answer, though, and both are easy to trip into. If the system infers a special-category attribute — ethnicity, sexual orientation, political opinion — to do its job, it becomes prohibited under Article 5(1)(g). And if age estimation feeds a social-scoring use, that is a different prohibition (Article 5(1)(c)). Get the design wrong and a low-obligation feature becomes a banned one.
This page draws the line precisely: why age estimation is usually outside Annex III, what you actually owe, and the two boundaries that turn it prohibited.
Why Age Estimation Is Usually Not Annex III High-Risk
The high-risk tier is a list. Annex III point 1 names three biometric uses that are high-risk: remote biometric identification (matching a face against many records to establish who someone is), biometric categorisation according to sensitive or protected attributes, and emotion recognition. Age estimation is none of these.
- It does not identify you. It does not match your face against a database to find your name; it estimates a number. That keeps it out of the remote-identification limb.
- It does not, by default, categorise you by a sensitive attribute. Apparent age is not on the Article 5(1)(g) protected list (race or ethnic origin, political opinions, trade-union membership, religious or philosophical beliefs, sex life or sexual orientation).
So plain facial age estimation for age-gating is not a named Annex III use. It is limited-risk as a regulatory tier: the obligation that actually attaches is transparency, not the full high-risk stack.
One nuance worth stating plainly: age estimation still processes biometric data (GDPR Article 9 special category), so data-protection law applies in full even though the AI Act high-risk tier does not. "Not Annex III high-risk" is not "unregulated".
What You Actually Owe: Article 50(3) and GDPR Article 9
Two duties carry the load for compliant age-estimation deployments.
Article 50(3) — disclosure of biometric processing. A deployer of a biometric categorisation or emotion-recognition system must inform the natural persons exposed to it. Apparent-age grouping is a form of biometric categorisation, so a clear notice at the point of capture — "we use facial analysis to estimate your age" — is the AI Act transparency duty. Crucially, Article 50, including 50(3), has applied since 2 August 2026 for everyone; it was not delayed by the high-risk deferral.
GDPR Article 9 — special-category data. A face image used to infer age is biometric data processed to estimate a characteristic of a person. You need a lawful basis under Article 9, a data-protection impact assessment under GDPR Article 35 where the processing is high-risk to rights, data minimisation (estimate and discard, do not store the image), and the usual transparency and rights machinery. Many age-estimation vendors are built precisely to return only an age band and delete the image — that design choice is what keeps the GDPR exposure proportionate.
A short scope note: this is age estimation from a face, not age verification against an ID document. Document-based verification is a different processing operation with its own data set, but it is likewise not an Annex III high-risk use on its own.
The Two Lines That Turn Age Estimation Prohibited
Here is where a benign feature becomes a banned practice. Both lines are about purpose and inference, not the underlying technology.
Line one — inferring a special-category attribute (Article 5(1)(g)). If the model, in order to gate access, deduces or infers race or ethnic origin, sexual orientation, religious belief, or any other Article 5(1)(g) attribute, the system is prohibited — full stop, since 2 February 2025. A vendor that "improves age accuracy" by inferring ethnicity has built a banned categoriser. The fix is architectural: the system must be demonstrably unable to infer a protected attribute.
Line two — social scoring (Article 5(1)(c)). If age estimation feeds a system that scores or classifies people over time based on social behaviour or inferred characteristics, leading to detrimental or unjustified treatment in unrelated contexts, that is the social-scoring prohibition — also live since 2 February 2025. Age-gating to block a minor from alcohol is fine; folding an age signal into a general "trustworthiness" score that follows a person across services is not.
| Use of age estimation | Tier | Basis |
|---|---|---|
| Estimate age band to gate adult content / restricted goods | Limited (transparency) | Outside Annex III; Article 50(3) + GDPR Art 9 |
| Demographic analytics on apparent age, anonymous and aggregate | Limited / minimal | Not a named Annex III use |
| Infer ethnicity or sexual orientation to "improve" gating | Prohibited | Article 5(1)(g) |
| Feed a cross-context "trust score" with detrimental effects | Prohibited | Article 5(1)(c) |
| Match the face against a database to identify the person | High-risk | Annex III point 1 (remote identification) |
Age Estimation vs Biometric Categorisation vs Identification
The cleanest way to avoid mis-tiering is to ask what the system outputs.
- Age estimation outputs an apparent age or band. No identity, no sensitive attribute. Limited-risk, transparency-only — unless it crosses into inferring a protected attribute (then prohibited).
- Biometric categorisation assigns a person to a group from their biometric data. If the group is a non-sensitive attribute (age band, broad demographic), it is not a named Annex III high-risk use — it sits at the transparency tier (Article 50(3)) with GDPR Article 9 on top. Biometric categorisation is high-risk under Annex III point 1 only where it sorts people by sensitive attributes; if it infers a sensitive attribute on the Article 5(1)(g) list, it is prohibited.
- Biometric identification establishes who a person is by matching against records. One-to-many remote identification is high-risk (Annex III point 1(a)); real-time remote identification in publicly accessible spaces for law enforcement is largely prohibited under Article 5(1)(h).
Age estimation sits at the gentle end of that spectrum — but it shares a body of technology with the categorisation case, which is why vendors must be explicit that their model neither identifies nor infers a sensitive trait.
Worked Example: VinoDirect S.L., an Online Wine Retailer
VinoDirect S.L., a Spanish direct-to-consumer wine retailer, ~60 employees, ~€9M annual turnover, must stop minors buying alcohol. At checkout it offers facial age estimation: the shopper takes a selfie, a third-party model returns an estimated age band, the image is discarded, and anyone estimated under the threshold is routed to document verification.
Classification: this is not an Annex III high-risk use. The system estimates age — it does not identify the shopper and does not infer any Article 5(1)(g) attribute. As a regulatory tier it is limited-risk. VinoDirect's duties are the Article 50(3) notice ("facial age estimation is used; your image is not stored") shown before capture, and full GDPR Article 9 compliance — lawful basis, a DPIA, immediate deletion of the image, and a non-biometric fallback for anyone who declines.
The trap VinoDirect must avoid is one procurement question away: if a vendor's model boosts accuracy by inferring ethnicity, the system flips to prohibited under Article 5(1)(g) — exposure of up to €35 million or 7% of worldwide turnover (Article 99(3)), with the SME cap under Article 99(6) taking the lower figure. VinoDirect's contract should warrant that the model infers age only, and cannot derive any protected attribute. With that warranty in place, the build is a transparency-and-data-protection exercise, not a high-risk programme.
Deadlines: Which Date Governs Age Estimation
Because age estimation is limited-risk, the high-risk deferral debate is mostly a distraction for it. The duty that binds is Article 50(3) transparency, which has applied since 2 August 2026 — for everyone, not delayed. The Article 5(1)(g) and Article 5(1)(c) prohibitions are already in force, since 2 February 2025, so the "don't infer a sensitive attribute, don't feed social scoring" line binds you now.
For completeness: the high-risk Annex III obligations (which would only matter if your system crossed into identification or sensitive-attribute categorisation) originally read 2 August 2026 in the statute; the Digital Omnibus on AI — Regulation (EU) 2026/1744, in force since 27 July 2026 — moved them to 2 December 2027. For a plain age-estimation deployment, the operative obligations are the Article 50(3) transparency duty, live since 2 August 2026, and the prohibitions.
How Confir Helps
Confir's rule-based engine tiers a biometric feature by what it outputs and infers. For age estimation it runs the Article 5(1)(g) sensitive-attribute test and the Article 5(1)(c) social-scoring test first — so a design that would slip into the prohibited zone surfaces as a stop finding — then confirms the limited-risk path and maps the Article 50(3) disclosure and the GDPR Article 9 cross-reference. The classification is deterministic and rule-based: the same intake yields the same finding, and the rule that fired is shown in plain language, so you can show a regulator why the system is transparency-tier rather than high-risk.
Related guides
- Biometric categorisation AI
- Biometric categorisation — glossary
- Article 50 transparency
- Voice recognition AI
Manage your EU AI Act compliance in one place
Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.
Start free trial →