Leaving this course?

Lesson 1 of 12

Three different questions, routinely conflated

"Age assurance" covers three operations that differ in what they establish, what they cost and whom they exclude. Using the words interchangeably is the first error, and it is usually made in the requirement rather than in the implementation.

There is now an international standard, and its title settles the vocabulary. ISO/IEC 27566-1:2025, Information security, cybersecurity and privacy protection — Age assurance systems — Part 1: Framework, published in December 2025, makes age assurance the umbrella and puts verification and estimation beneath it. The standard's text is sold rather than published, so this course does not quote it; the naming it fixes is visible in its title. The statutory definitions are unusually clean. The UK Online Safety Act 2023 defines two of them in section 230. Age verification "means any measure designed to verify the exact age of users of a regulated service." Age estimation "means any measure designed to estimate the age or age-range of users of a regulated service."

Exact age against an age range. A document or an authoritative record gives the first; a model looking at a face or at behaviour gives the second.

And the Act excludes the thing most services actually do. Section 230(4): "A measure which requires a user to self-declare their age (without more) is not to be regarded as age verification or age estimation."

A date-of-birth box is not age assurance in law. That sentence disposes of a large proportion of what is deployed on the internet today, and it is worth quoting to anybody who believes their registration form already complies with something.

Age inference is the third, and it is not defined in the Act. Inferring an age band from signals — the account's history, the device, the content consumed, a payment instrument that requires adulthood to hold. It is the least intrusive and the least reliable, and because it is undefined in that legislation you cannot rely on a statutory label for it.

What distinguishes them in substance. The European Commission's guidelines on protecting minors put it in one line: "The main difference between age estimation and age verification measures is the level of accuracy."

Those guidelines also set out the three-way split explicitly — self-declaration, age estimation and age verification as separate categories — which is the vocabulary to use when writing a requirement, because it forces the author to say which one they mean.

Why the distinction decides the design. The three fail differently:

  • Verification fails by exclusion. A person without the document or the record cannot pass, however old they are.
  • Estimation fails by error, and the error is not evenly distributed. Lesson four shows that the false positive rate for a 25-year threshold rises by an order of magnitude between a fourteen-year-old and a twenty-year-old.
  • Inference fails silently. A wrong inference produces no signal at all, because there is nothing to check it against.

And one framing to carry through the course. The question is almost never "how old is this person". It is "is this person over a threshold", which is a smaller question with a smaller answer — and the whole privacy architecture in lesson seven exists because the smaller answer is the one you are entitled to.

Lesson 2 of 12

Why this arrived all at once

Age assurance moved from a niche compliance topic to a mainstream engineering requirement in about three years, because four separate regimes arrived in quick succession. Knowing which applies to you is most of the work.

The United Kingdom: duties with dates. The Online Safety Act 2023 creates two relevant duties, and their commencement is recorded in the legislation itself rather than in commentary.

Section 12, the safety duties protecting children, shows commencement on 10 January 2024 by statutory instrument. Its operative provision: "The duty set out in subsection (3)(a) requires a provider to use age verification or age estimation (or both)" to prevent children encountering the most harmful categories of content, with an exception where the terms of service prohibit that content altogether.

Section 81, on a provider's own pornographic content, came into force on 17 January 2025, again by statutory instrument. It imposes "A duty to ensure, by the use of age verification or age estimation (or both), that children are not normally able to encounter content that is regulated provider pornographic content in relation to the service." It also requires a written record and "A duty to summarise the written record in a publicly available statement."

The European Union: an obligation without a method. The Digital Services Act, Article 28(1): "Providers of online platforms accessible to minors shall put in place appropriate and proportionate measures to ensure a high level of privacy, safety, and security of minors, on their service."

Appropriate and proportionate, with no method named. The Commission's 2025 guidelines under Article 28(4) fill that in, and lessons seven to nine draw on them.

Alongside it, the revised Audiovisual Media Services Directive has required age measures for longer: "Such measures may include selecting the time of the broadcast, age verification tools or other technical measures", with the rule that "The most harmful content, such as gratuitous violence and pornography, shall be subject to the strictest measures."

Australia: a minimum age, by statute. The Online Safety Amendment (Social Media Minimum Age) Act 2024 was assented on 10 December 2024. It defines an "age-restricted user" as "an Australian child who has not reached 16 years", and obliges a provider of an age-restricted social media platform to "take reasonable steps to prevent age-restricted users having accounts", with a civil penalty of 30,000 penalty units.

The United States: a rule with a compliance date. The Children's Online Privacy Protection Rule amendments were published in the Federal Register on 22 April 2025. The amended rule "is effective June 23, 2025", and "regulated entities have until April 22, 2026 to comply", with limited exceptions. It defines a child as "an individual under the age of 13."

And the thread through all four. Each requires an outcome and leaves the method to the provider, which is why the rest of this course is about choosing a method and being able to defend it, rather than about implementing a prescribed one.

Lesson 3 of 12

"Highly effective", and what this course cannot tell you

The UK standard is a phrase rather than a number, and it appears in both duties in identical words.

The statutory test. Section 12(6): "the age verification or age estimation must be of such a kind, and used in such a way, that it is highly effective at correctly determining whether or not a particular user is a child." Section 81(3) repeats it verbatim for provider pornographic content.

Three parts are doing work in that sentence. Of such a kind — the method itself must be capable. Used in such a way — a capable method deployed badly does not satisfy it, so configuration and thresholds are in scope. And correctly determining whether or not a particular user is a child — the test is about a binary outcome for an individual, not an aggregate accuracy figure for a population.

What this course does not state, and why. The regulator has published guidance on what it considers highly effective, including criteria and a list of methods it regards as capable of meeting the test. Every attempt to retrieve that guidance for this course was refused by the regulator's own website, which blocks automated access.

So this course does not state those criteria, that list, or the commonly cited date on which the children's safety duties came into force — because the only sources available for them were secondary. The commencement dates in the previous lesson come from the legislation's own commencement notes, which is why they are given and the other date is not.

What follows for you, practically. Read the regulator's guidance directly. It is the document a supervisor will hold you to, and no course should stand between you and it. What this course can give you is everything the guidance will assume you already understand: how the methods work, what their error rates actually look like, how to build one that does not hoard personal data, and who each method excludes.

And one inference that is safe to draw from the statutory words alone. "Highly effective at correctly determining whether or not a particular user is a child" cannot be satisfied by a method whose error rate is unknown to you. If you cannot say how often your method is wrong, in which direction, and for whom, you are not in a position to assert that it is highly effective — whatever your vendor's marketing says.

That is the argument for lessons four and five, which are about measurement, and it holds independently of any regulator's criteria. A provider who has measured its own error rates on its own population is in a defensible position under any of the four regimes in the previous lesson. One who has not is relying on a supplier's claim about somebody else's population.

Lesson 4 of 12

Face-based age estimation: the actual numbers

Estimating age from a photograph is the method that has driven most of the recent deployment, because it needs no document and no authoritative record. Its accuracy is measured publicly, and the figures are more interesting than either its advocates or its critics suggest.

The evaluation. NIST runs a continuing evaluation of age estimation and verification algorithms and publishes the results. Its first public report, in May 2024, records the trajectory: "Using a common visa database, best mean absolute error (MAE) has reduced from 4.3 to 3.1 years."

For context, the 4.3-year figure comes from NIST's own 2014 evaluation, which found "the chance of accurately estimating the age of a person within five years of their actual age over an ethnically-homogeneous database of 6 million images is 67%, with an MAE of 4.3 years." A decade of improvement moved the best mean error from about four and a third years to about three.

Three years of average error, and why that is not the headline. An average error of three years sounds disqualifying for an eighteen-year threshold and is not, because of how the method is deployed.

The challenge policy is the trick. NIST describes it: "For restricted-age applications such as alcohol purchase, a Challenge-T policy accepts people with age estimated at or above T but requires additional age assurance checks on anyone assessed to have age below T."

Set the challenge threshold well above the legal one — 25 for an 18 limit — and the estimator becomes a filter rather than a decision. Anyone it reads as clearly older passes; everybody else is routed to a stronger check. The population routed onward is small, which is the point: the expensive or intrusive method is used on a fraction of users.

And the number that shapes the design. The same report: "Challenge-25 false positive rates increase by an order-of-magnitude as subjects age from 14 through 20 — for example for the roc-000 algorithm FPR at age 20 (0.295) is almost fifteen times larger than at age 14 (0.02)."

Read that carefully, because it is the most useful fact in this lesson. The method is good at identifying children who are clearly children and poor at the boundary. A fourteen-year-old is rarely passed as over 25; a twenty-year-old frequently is. For protecting young children the method is strong. For distinguishing a seventeen-year-old from an eighteen-year-old it is weak, and no threshold setting fixes that.

One more figure, for the opposite end. Accuracy varies enormously with age: "mean absolute error in children under age 1 is 0.34 years (i.e. about four months) versus 3.5 years for a thirty year old." Estimation is far more precise on the very young, which is intuitive once stated and is the reason age-band checks for young children behave better than adult-boundary checks.

Lesson 5 of 12

Why the headline metric is the wrong one

The figure every vendor quotes is mean absolute error, and the body that publishes it says it is the wrong measure for the job most buyers are doing.

The warning, from the evaluation itself. NIST's report states that mean absolute error "is not an appropriate metric for age verification (AV) tasks".

The reason is structural. Mean absolute error averages how wrong the estimate is across the whole population. An age verification task asks a different question — is this person above a threshold — and only errors near that threshold matter. An algorithm with an excellent average, achieved by being very accurate on children and adults far from the boundary, can be poor at exactly the decision you are making.

So a procurement that compares two products on mean absolute error is comparing them on a measure that does not describe the task.

What to ask for instead. The error rates at your threshold, for the age bands either side of it, stated separately:

  • The proportion of under-age users the method passes, at your configured threshold.
  • The proportion of over-age users it refuses, at the same setting.
  • Both figures broken down by age band, because lesson four showed they move by an order of magnitude across a six-year span.
  • And both broken down by the demographic groups in your actual population.

A correction worth knowing about, because it affects published figures. The evaluation is a living document, revised repeatedly. Its release notes record that in April 2026, "The weights used in the ineffectiveness formula eq. 14 were calculated incorrectly", with the consequence that "The old formula yielded higher weights for 13 year olds than for 17 years olds."

That matters for anyone who quoted the composite "ineffectiveness" figures, or the rankings derived from them, between the first publication and that correction. The raw error rates in the original report are unaffected; the weighted composite and the ordering are.

So cite the edition. The May 2024 report and the continuously updated draft revision are different documents with the same number, and by late 2026 the revision was on its twenty-fourth update. A reference that says only the report number does not identify which figures were read. This is the same discipline the cryptography course reaches about standards: check the status of a document at the moment you rely on it.

And what the evaluation defines that is worth adopting. The programme page sets out the metric that actually describes a child-protection task: for the 13 to 16 band, "False Negative Rates (FNR) are proportions of subjects aged 13 to 16 whose age is incorrectly estimated as 8 to 12."

That is a measure of a specific confusion between two specific bands, which is what a requirement should name. It also notes the dependency that dominates everything: "False negatives are strongly dependent on image quality, and poor photography of a face can induce a demographic effect."

Poor capture does not merely add noise. It creates apparent demographic differences that are properties of the photography rather than of the people — which means a service with a bad camera flow has a fairness problem it will misattribute to the algorithm.

Lesson 6 of 12

The other routes, and who each one shuts out

Estimation is one method among several, and a design that offers only one has decided who it will exclude without saying so.

Document checks. An identity document, read and validated, with the date of birth taken from it. This is age verification in the statutory sense: exact age from an authoritative artefact. It is the strongest method and the most exclusionary — it requires the person to hold a document, to have it to hand, and to be willing to show a government credential to a commercial service in order to read a web page.

The disproportion is the objection. Producing a passport to enter a bar is accepted; producing one to visit a website is a different exchange, and the data protection authorities in lesson nine treat it as one.

Checks against an authoritative record. Where law permits, confirming a date of birth against a population register or a credit record. Accurate, invisible to the user, and dependent on the person being in the record — which excludes the recently arrived, the young with no credit history, and in many countries a substantial share of the population.

A reusable credential. The architecture the wallets course describes, used for its narrowest case: proving an age band without disclosing anything else. The European identity framework names the enabling property directly — "Selective disclosure is a concept empowering the owner of data to disclose only certain parts of a larger data set, in order for the receiving entity to obtain only such information as is necessary for the provision of a service requested by a user" — and requires that wallets "technically enable the selective disclosure of attributes to relying parties".

It also contemplates pseudonyms: wallets "should include a functionality to generate user-chosen and managed pseudonyms, to authenticate when accessing online services." An "over 18" token with a user-managed pseudonym is the cleanest answer available to the whole problem, and its limitation is adoption, not design.

Inference from account signals. History, device, payment instrument, behaviour. Cheap and unverifiable, and it carries a particular hazard: profiling a user's behaviour to guess their age is itself processing of personal data, undertaken to avoid processing personal data. Lesson nine returns to that.

Vouching by an adult. Parental consent flows, which the amended US rule treats as a distinct problem with enumerated methods — including one worth noting for what it demands: "Having a parent submit a government-issued photographic identification that is verified to be authentic and is compared against an image of the parent's face taken with a phone camera or webcam using facial recognition technology", with prompt deletion of the identification and images required afterwards.

The design conclusion. No single method is adequate, and the European guidelines state the remedy as an expectation rather than a suggestion: a platform "should offer on its platform more than one age assurance method, to provide the user with a choice between methods". Offer at least two that do not share a dependency, and measure which one people actually use.

Lesson 7 of 12

The double-blind architecture

The reason age assurance is feared is not the check. It is the record the check leaves behind: a commercial service that now knows your date of birth, holds a photograph of your face, or has seen your passport, in order to let you read something.

That outcome is avoidable by architecture, and the design has a name.

What it separates. The European Commission's guidelines describe the property precisely: "A double-blind method ensures that (i) the online platform does not receive additional means to identify the user and, instead only receives information allowing it to confirm whether they meet the required age threshold" — and the second limb completes it, so that the age verification provider does not learn which service the proof was used for.

Two parties, each blinded to a different thing. The service learns a yes or no and gains no new identifier. The provider learns who the person is and not where they went.

The same design, described earlier. The French data protection authority set out the separation in 2022: "the entity providing the proof of age knows the identity of the Internet user but does not know which site the latter is visiting." That page predates the Commission's guidelines and the shared approach is the point — this is a settled design rather than a novel one.

And a principled version from Spain. The Spanish authority's technical note states two of its principles in terms worth lifting into a requirement document directly. "Principle 3: Evidence of exceeding the required age threshold should be anonymous to Internet service providers and third parties." And "Principle 7: Age verification should not make it possible to link a person's activity across different Internet services."

Principle 7 is the one most designs fail. A token that is the same token everywhere is a tracking identifier that happens to assert an age, and it makes the age-assurance provider into the most comprehensive observer of a person's browsing in existence. Everything the wallets course says about unlinkability applies here, for the same reasons.

What the service should end up holding. The Commission is explicit about the limit: "Age verification should not entitle providers of online platforms to store personal data beyond information about the user's age group."

An age group. Not a date of birth, not an image, not a document number, not an identifier that works across sites.

And the honest caveat on all of it. The French authority's assessment of the market at the time was blunt: "At present, all the solutions proposed can easily be circumvented." A determined teenager with access to an adult's device, document or token defeats any of this. That is not an argument against building it well; it is an argument against claiming more than it delivers, and against imposing heavy data collection on the entire adult population for a control that the motivated minority will route around.

Lesson 8 of 12

Minimisation, and what you are allowed to keep

Every age check generates data that was not needed a moment earlier, and the discipline is to let as little of it persist as possible.

The governing principle, in its own words. The GDPR requires personal data to be "adequate, relevant and limited to what is necessary in relation to the purposes for which they are processed ('data minimisation')". That is the clause every rule in this lesson derives from.

Collecting the date of birth when you need a threshold is a breach of it. The Spanish authority's technical note says so directly: "Collecting the date of birth or precise age of Internet users when not required infringes the minimisation principle."

Worth dwelling on, because a date-of-birth field is the default implementation everywhere. If the decision you need is over or under a line, the exact date is more data than the purpose requires — and as lesson one established, a self-declared date is not age assurance in any case. So the commonest implementation collects data it may not keep in order to achieve nothing.

What the European Data Protection Board says the service needs. "For example, the service provider may only need to know whether the user is over or under an age threshold." A single bit, in the general case.

And when the data may persist. The Irish Data Protection Commission addresses the retention question head-on, first clearing an objection — "The DPC does not consider the principle of data minimisation to be an obstacle to age verification" — and then setting the rule: "personal data collected, which provides the basis for the age verification process to be undertaken, is deleted once the appropriate level of confidence as to user's age has been attained."

Deleted once the confidence is attained. Not retained for audit, not retained in case of a dispute, not retained because the logging framework captures everything by default. The evidence that justified the conclusion goes; the conclusion stays.

That is a harder engineering requirement than it sounds, because it has to hold in the places the eKYC and data protection courses both identify as the forgotten stores: the image buffer, the request log, the vendor's own retention, the analytics pipeline and the backup.

The one thing worth keeping. A record that a check was performed, by which method, with what outcome, and when — enough to show a supervisor that the duty was discharged, and containing no personal data beyond the age band. The UK Act requires a written record for the provider pornographic content duty and a public summary of it, so for that duty the record is not optional.

And the retention rule in a jurisdiction that states it generally. Kenya's Data Protection Act: "A data controller or data processor shall retain personal data only as long as may be reasonably necessary to satisfy the purpose for which it is processed". The purpose of an age check is satisfied at the moment the answer is known.

Lesson 9 of 12

Proportionality: when you may not check at all

The question most age assurance projects never ask is whether the check is lawful in the first place. Checking is itself processing, and processing needs a justification.

The clearest statement of the limit. The European Data Protection Board: "a service provider processing personal data to check the age of all their users when accessing all their content or services, even when the content or services are suitable for all audiences and devoid of any risk, would not pass the necessity and proportionality tests".

A blanket check across an entire service, including its harmless parts, fails. That is a conclusion about design, not about implementation quality — a perfectly built double-blind check applied where it is not needed is still disproportionate.

And the legislation says the same from the other direction. The Digital Services Act, Article 28(3): "Compliance with the obligations set out in this Article shall not oblige providers of online platforms to process additional personal data in order to assess whether the recipient of the service is a minor."

The duty to protect minors is not a mandate to identify users. A provider who reads Article 28(1) as requiring age verification everywhere has read past paragraph 3.

So the design starts with a map, not a method. Which parts of the service carry a risk to a child, and what the risk is. The answer is usually that a minority of the surface needs a gate, and that the gate belongs at the point of the risky action rather than at the front door. A platform that checks age at registration and never again has imposed the cost on everybody and protected nothing at the moment of exposure.

The judgement is yours and it is not transferable. The Irish Data Protection Commission puts it plainly: "What may be considered a suitable approach for one organisation may be entirely unsuitable for another."

Which means a supplier's compliance claim is not your compliance. The assessment of what is proportionate for your service, your content and your users is an assessment you have to make and record, and nobody can make it for you.

What goes in that record. Which content or actions were assessed as carrying risk, and why. Which were assessed as not, and why. What method was selected for each gate and what alternatives were considered. What data the method processes and for how long. And the date, because the assessment is about a service that changes.

One practical consequence, worth naming because it is counter-intuitive. The strongest argument against over-checking is not privacy; it is the error analysis in lessons four and five. Every check has a false rejection rate. Applied to the whole user base rather than to the risky surface, that rate turns into a volume of excluded adults large enough to matter commercially — which means proportionality and self-interest point the same way, and the privacy argument is the one you can also take to a finance committee.

Lesson 10 of 12

The thresholds are not one number

"Age assurance" implies a single line to be above or below. There are several, they differ by jurisdiction and by purpose, and a system built around one of them will fail the others.

The data protection consent age. The GDPR sets a default and a floor: processing a child's personal data on consent is lawful "where the child is at least 16 years old", and "Member States may provide by law for a lower age for those purposes provided that such lower age is not below 13 years."

So within one regime the relevant age may be anywhere from 13 to 16 depending on the member state. A single European deployment is dealing with a range, not a number.

Below that age the regime shifts to the parent: "The controller shall make reasonable efforts to verify in such cases that consent is given or authorised by the holder of parental responsibility over the child, taking into consideration available technology."

Note the qualifier — reasonable efforts, taking into consideration available technology. That is a standard that moves as the technology improves, which means a method that was reasonable five years ago may not be now.

The US children's privacy line. The amended rule is unambiguous: "Child means an individual under the age of 13." A different number from the European default, with a different regime attached, and a compliance date of 22 April 2026.

The Australian social media minimum age. "age-restricted user means an Australian child who has not reached 16 years" — a prohibition on holding an account rather than a consent threshold, enforced with a civil penalty on the platform.

And the UK content duties use "child". The Online Safety Act's test is whether a user is a child, which is a single line at eighteen for the content duties in lesson two — a different question again from consent or from account-holding.

What this means for a system that operates in more than one place. Four observations, each of which has caught a real deployment:

  • The threshold is configuration, not code. Anything that hard-codes eighteen will be rewritten.
  • The consequence of being below it differs. Refuse the account, obtain parental consent, restrict the content, or restrict the features — these are four different flows, not one.
  • The evidence required differs with the consequence. A feature restriction can rest on an estimate; refusing an account on a statutory prohibition probably cannot.
  • And the bands you must distinguish differ. Separating under-13 from over-13 is a different measurement problem from separating under-18 from over-18, and lesson four showed the second is much harder because estimation degrades at the adult boundary and is precise on the young.

The design conclusion. Build a service that answers "is this user above threshold T, to confidence C" for a configurable T, and record which T was applied. A system that answers only "is this user an adult" cannot be deployed into a second jurisdiction without being rebuilt.

Lesson 11 of 12

The adult who cannot pass

Every age assurance method refuses some people who are entitled to proceed, and because the refused population is not the one designing the system, the refusal is usually invisible until somebody complains publicly.

Who they are. An adult with no identity document, or none to hand. A person whose appearance reads younger than their age, which lesson four showed is a measurable and unevenly distributed effect. Someone with no credit record, no bank account and no presence in a population register. A person sharing a device with a household. Someone whose disability prevents a gesture, a capture or a reading. And anyone for whom showing a government document to a commercial service is not an acceptable exchange, which is a legitimate position rather than an obstruction.

The obligation to offer another route. The European Data Protection Board states it as an expectation with the reasons attached: "for example because they do not own a suitable identity document or mobile phone or because of a disability, alternative methods for age assurance should be made available, when reasonably possible".

And the Commission's guidelines make the plural a design requirement rather than a courtesy — a platform "should offer on its platform more than one age assurance method, to provide the user with a choice between methods".

Choice between methods. Not a documented exception process buried in a help article, which is what most services provide and what nobody finds.

The criteria to design against. The same guidelines set out five properties for an age assurance method: accuracy, reliability, robustness, non-intrusiveness and non-discrimination. The last two are the ones that fail quietly. Non-intrusiveness is a judgement about whether the data demanded is proportionate to the content being gated. Non-discrimination is the error analysis from lesson five, applied to the groups in your own population rather than to a vendor's test set.

And the point about self-declaration, which bears repeating here. The Board observes that "robustness has little meaning in the context of the self-declaration of an age-related attribute". A date-of-birth box is not a robust method and cannot be made one, so it is not the inclusive alternative to a strong check. It is the absence of a check.

The inclusive alternative has to be a genuine second method — a different route to the same assurance — not a weaker one offered to whoever asks.

How to know whether the alternative works. Measure it. How many users were refused by the primary method, how many then used the alternative, and how many simply left. The third number is the one nobody collects, and it is the only measure of exclusion you will get, because an excluded user does not raise a ticket.

The sentence to design against. A method that refuses a meaningful share of legitimate adults has not made the service safer for children; it has made it unavailable to adults, and the children who were motivated to get past it have used somebody else's device.

Lesson 12 of 12

Kenya, and a contradiction worth teaching

Kenya's framework states the obligation generally rather than by sector, and reading it alongside the material in lesson one produces a disagreement that is more instructive than either source alone.

The statutory duty. The Data Protection Act, 2019, section 33(2): "A data controller or data processor shall incorporate appropriate mechanisms for age verification and consent in order to process personal data of a child."

A general obligation on anyone processing a child's personal data, not limited to social media, pornography or video sharing. Any service that might process a child's data is in scope, which is a broader reach than the sectoral duties in lesson two.

And it tells you how to decide what is appropriate. Section 33(3) sets out the factors, including "proportion of such personal data likely to be that of a child" and "possibility of harm to a child arising out of processing of personal data".

That is a risk-based test, and it maps closely onto the proportionality analysis in lesson nine: how much of your data is likely to be children's, and what harm could follow. A service with almost no child users and no harmful content is being told something different from a service aimed at teenagers.

Section 33(1) also requires that processing protects and advances the rights and best interests of the child, and the Act's general retention rule from lesson eight applies: personal data retained "only as long as may be reasonably necessary to satisfy the purpose for which it is processed".

Now the contradiction. The Office of the Data Protection Commissioner's guidance for the education sector explains the mechanisms in practical terms: "Age verification helps to ensure that children do not access inappropriate content or services that could be harmful to them." And it gives examples: "Examples of age verification methods include requiring users to enter their date of birth, using identity documents, or conducting age checks with a third-party service."

Requiring users to enter their date of birth, listed as an age verification method.

Compare section 230(4) of the UK Act — "A measure which requires a user to self-declare their age (without more) is not to be regarded as age verification or age estimation" — and the European Commission's guidelines, which treat self-declaration as a separate category from estimation and verification precisely because it establishes nothing.

What to do with that. Not dismiss the guidance. It is the regulator's published view in its own jurisdiction, and a Kenyan controller follows Kenyan law. The useful conclusion is narrower and more valuable:

"Age verification" is not one definition across jurisdictions. A control that satisfies a regulator in one place is explicitly excluded by statute in another. So a multi-jurisdiction service cannot implement "age verification" as a single feature against a single definition — it has to implement a method, know what that method actually establishes, and map it to each regime's test separately.

And the practical reading for a Kenyan service with international users. Build to the stronger definition. A method that satisfies the statutory test in lesson three will satisfy a guidance note that would also have accepted a date-of-birth box. The reverse is not true, and a service that built to the weaker definition will rebuild.

Knowledge check

Fifty questions

Answer all fifty questions, then check your answers. You need 45 out of 50 to pass and receive a certificate. If you score less, you will see which answers were right and wrong, and then go through the course again before you retake the check. Your answers, progress and times are kept only in this browser.

Standards and frameworks covered

The published frameworks this course teaches toward. Saolix is not affiliated with, endorsed by or accredited by any of these bodies; naming a standard is a citation, not a claim about its author.

Sources

The official documents this course relies on. Laws and guidance change, so check the current version.

  1. ISO/IEC 27566-1:2025, Information security, cybersecurity and privacy protection — Age assurance systems — Part 1: Framework · International Organization for Standardization and International Electrotechnical Commission
  2. NIST IR 8525: Face Analysis Technology Evaluation, Age Estimation and Verification (May 2024) · US National Institute of Standards and Technology
  3. NIST IR 8525, continuously updated revision (includes updates as of November 2026) · US National Institute of Standards and Technology
  4. Face Analysis Technology Evaluation: Age Estimation and Verification (programme page and metric definitions) · US National Institute of Standards and Technology
  5. NIST IR 7995: Performance of Automated Age Estimation Algorithms (March 2014, superseded; historical baseline) · US National Institute of Standards and Technology
  6. Online Safety Act 2023, section 230: age verification and age estimation · UK Parliament, via legislation.gov.uk
  7. Online Safety Act 2023, section 12: safety duties protecting children · UK Parliament, via legislation.gov.uk
  8. Online Safety Act 2023, section 81: duties about regulated provider pornographic content · UK Parliament, via legislation.gov.uk
  9. Data Protection Act 2018, section 123: the statutory basis of the age-appropriate design code · UK Parliament, via legislation.gov.uk
  10. Regulation (EU) 2016/679 (GDPR), Articles 5(1)(c) and 8 · Official Journal of the European Union, via EUR-Lex
  11. Regulation (EU) 2022/2065 (Digital Services Act), Article 28: online protection of minors · Official Journal of the European Union, via EUR-Lex
  12. Guidelines on measures to ensure a high level of privacy, safety and security for minors online (OJ C/2025/5519) · European Commission
  13. Statement 1/2025 on Age Assurance (adopted 11 February 2025) · European Data Protection Board
  14. Online age verification: balancing privacy and the protection of minors (September 2022) · Commission nationale de l'informatique et des libertes (CNIL), France
  15. A safe Internet by default for children and the role of age verification (October 2024) · Agencia Espanola de Proteccion de Datos (AEPD), Spain
  16. Fundamentals for a Child-Oriented Approach to Data Processing (December 2021) · Data Protection Commission, Ireland
  17. Regulation (EU) 2024/1183 (European Digital Identity Framework), recitals 22 and 59 · Official Journal of the European Union, via EUR-Lex
  18. Directive (EU) 2018/1808 (revised Audiovisual Media Services Directive), Articles 6a and 28b · Official Journal of the European Union, via EUR-Lex
  19. Children's Online Privacy Protection Rule, final rule amendments (90 FR 16918, 22 April 2025) · US Federal Trade Commission, in the Federal Register
  20. Online Safety Amendment (Social Media Minimum Age) Act 2024 (No. 127, 2024) · Parliament of Australia, Federal Register of Legislation
  21. Data Protection Act, 2019 (No. 24 of 2019), sections 33 and 39 · Office of the Data Protection Commissioner, Kenya
  22. Guidance Note for the Education Sector (December 2023) · Office of the Data Protection Commissioner, Kenya