Explainer: Age Inference

By now, most people are familiar with age verification, and facial age estimation has hit the mainstream in the EU and UK, as major websites implement tools to comply with new online safety regulations.

Age verification requires initial confirmation of a birth date, usually from an identity document, which is matched against face biometrics. Age estimation uses machine learning algorithms to analyze facial geometry and estimate an individual’s age within a small range.

While these two approaches have garnered the most attention in discussions about age assurance (the blanket term for systems that provide age checks), another term has emerged in recent documentation: age inference. The category gets its own volume in the newly published final report from the Age Assurance technology Trial (AATT) commissioned by the government of Australia.

The AATT’s glossary defines age inference as an age assurance method “based on verified information which indirectly implies that an individual is over or under a certain age or within an age range.” Alternately, “a method of determining an individual’s likely age or age range based on verifiable contextual, behavioral, transactional or environmental signals, rather than biometric data or identity documents.”

These signals may include facts such as enrollment in schools or voter logs, financial transactions, content settings, service usage, the age of a user’s accounts, or participation in age-specific activities.

How these factual signals are collected and bound to an individual varies by provider. Examples of inference methods used by AATT participants include email domain recognition, session metadata analysis, interaction patterns, credit eligibility and content engagement. Contextual references could include documents, records, images, online activities, or “any other authoritative and reliable source.”

YouTube recently introduced an age inference system that guesses a person’s age based on how they interact on the platform – what they watch, what they go back to, and so on. This variety of session-based age inference can also analyze interaction patterns like text input or navigation speed.

Some models use language analysis. If a user is engaged in passionate discourse about, say, Skibidi Toilet, they are probably not over 25.

However, the approach is anchored in common institutional norms and laws. If a user has an accessible student record, they can be assumed to be under 16. If they have a driver’s license, they can be assumed to be 16 or over. If their email is on a list for a whisky company, they’re more likely to be over legal drinking age. If they opened a bank account in 1993, they’re probably older than that.

Inferences will correspond with legal thresholds in particular jurisdictions. In Australia, one must be 21 to get a commercial pilot’s license; hence, a pilot’s license implies a user that is 21 or over.

Age inference is covered by ISO/IEC FDIS 27566-1, which, according to AATT, “treats age inference as a versatile and powerful age assurance tool, capable of supporting high-quality age-related decisions – provided it is implemented with clear logic, accountable governance and appropriate safeguards.”

Age inference providers listed in the AATT report include Yoti, Luciditi, Equifax, FrankieOne, IDVerse, MyMahi, Privo and VerifyMy.

Article Topics

age inference  |  age verification  |  biometric age estimation  |  digital identity  |  facial age estimation (FAE)  |  ISO 27566-1