Proving How Old You Are

54.0 What this chapter gives you#

1. You will be able to define age verification, age estimation, age inference and age assurance exactly as the law and the standards use them, and cite the statute section that fixes two of them.
2. You will be able to describe the five families of method, say what each checks, and name which a regulator accepts as capable of being highly effective.
3. You will be able to read a facial age estimation accuracy claim using mean absolute error, mean error, accuracy within three years and the challenge threshold, and say why one headline number tells you little.
4. You will be able to explain the buffer, or challenge age, approach with real numbers, and calculate what it costs a genuine twenty-four-year-old.
5. You will be able to state which duties the Online Safety Act 2023 imposes, on whom and from which dates, and give Ofcom’s four criteria for highly effective age assurance.
6. You will be able to summarize what the United States Supreme Court decided in Free Speech Coalition v Paxton, what standard of review it applied, and where American law is still split.
7. You will be able to place the European approach across the Audiovisual Media Services Directive, Article 28 of the Digital Services Act, the July 2025 blueprint and the April 2026 Recommendation.
8. You will be able to draw the double-blind pattern, explain a tokenised age credential, and say what a zero-knowledge proof of age adds and what it does not.
9. You will be able to quote real circumvention data, including virtual private network use before and after the United Kingdom rules, and say what it does and does not prove.
10. You will be able to design an age assurance flow for a real service, price it, and defend it to a regulator, a privacy officer and a furious adult user.

Almost every country has rules about age. You must be a certain age to buy alcohol, to drive, to marry, to vote, to gamble, to see certain films. For most of history those rules were enforced by a person looking at another person, in a shop or at a door, and making a judgement. That system was crude, biased and easy to cheat, and it worked well enough because the transaction happened in a physical place with a human being who could be held responsible.

The internet removed the human being and the place. For roughly thirty years the answer was a box saying “I am over 18” with a button next to it, and everybody involved knew it was theatre. Between 2022 and 2026 that stopped being acceptable in law across a large part of the world, and the change came very fast. As of August 2026 the United Kingdom requires age checks on pornography and on the riskiest content for children, the United States Supreme Court has held that a state may require proof of age before showing sexual material to adults, more than half the American states have such laws, the European Commission has published a reference implementation of an age proof and asked every member state to have one by the end of 2026, and Australia has banned social media accounts for under-sixteens outright, with the United Kingdom announcing it will do the same.

This is the fastest-moving area of identity regulation in the world and the hardest to get right, because it contains a genuine conflict rather than a solvable engineering problem. The state wants a check that cannot be fooled. That means strong evidence, which usually means identity documents, which means somebody now knows exactly who visited that website. The technical art of age assurance is the art of getting the first thing without the third.

That is the thesis of this chapter. We will build the idea up from a bouncer at a door, break the analogy where it fails, and then go through the specifications, statutes, judgments, accuracy figures and circumvention data with a date on every one. Chapter 53 covers the signed credentials a person carries in a wallet, and chapter 55 covers the data protection law behind all of this, so we touch both only where they bear directly on age.

The plain version#

The bouncer and the birthday#

Picture a nightclub with a person on the door whose job is to keep out anyone under eighteen. They have three ways to do it.

The first is to ask for a document. You hand over a passport, they read the date of birth, do the arithmetic, and look at the photograph and then at your face to check the document is yours. That gives them your exact age to the day.

The second is to look at you and guess. Experienced door staff are surprisingly good at this. If you look forty, they wave you through. If you look eighteen, they ask for the document. The guess decides whether the expensive check is necessary.

The third is to trust somebody else’s work. If you walk in with a wristband from the bar next door, and the bouncer knows the bar next door checks properly, the wristband is enough. Nobody looks at a document, because somebody already did.

Those three moves are the whole of age assurance. Everything in this chapter is a version of one of them, done by machines, millions of times a day, under laws that did not exist five years ago. The difference between the first and the second matters enough that the law writes it down.

Two ways to answer, and one word for both#

Checking a document tells you a date. That is age verification: measuring somebody’s actual age against a real record of when they were born. The output is a fact, subject to the record being right and the document being theirs.

Looking at a face tells you a guess. That is age estimation: a likely age, or age range, from something about the person rather than a record of their birth. The output is a probability dressed up as a number.

A third move is neither. If somebody holds a credit card, and a bank will only give a credit card to an adult, you know they are an adult without seeing a date of birth or a face. You have inferred the age from something already checked by somebody with a reason to check it properly. That is age inference.

One word covers all three, because a real system uses more than one: age assurance. It is the umbrella. If a product does age assurance, you have been told the category and not the method, and you should ask which of the three it is.

Five ways a website can find out how old you are#

Strip away the branding and there are five families, each trading the same two things, how sure the answer is against how much the checker learns about you.

The first is the document. You photograph your passport or driving licence, then take a short video of your face so the system can check that a real, present human is holding it rather than somebody waving a printed photograph at a camera. The system reads the date of birth, checks the document’s security features, and compares the face in the document with the face in the video. Strongest evidence available, and the most intrusive.

The second is a record somebody else keeps. Your bank knows how old you are, because it had to check before opening the account. Your mobile network knows, because adult content filters on British networks are on by default until the account holder proves their age. Any of these can be asked one narrow question, whether this person is over eighteen, and answer from a record they already hold, with no document changing hands.

The third is your face. A model trained on millions of photographs of people whose real ages are known looks at your selfie and outputs a number. No document, no name, and in the better implementations the image is deleted within seconds. Fast, cheap, and a guess.

The fourth is somebody vouching for you. A parent confirms that a child is a child; an adult with a verified account confirms the person beside them is over eighteen. Common in family account setups, rare as a primary check for adult content, because the obvious attack is to find a friendly adult.

The fifth is a hard identifier: a government-issued number tied to a state record of your birth. Where a country has one and uses it, the check is trivially accurate and maximally identifying, because it is the same number used for everything else about you.

The trick of asking for more than you need#

Here is the cleverest idea in the whole subject, and it is easy to explain.

Suppose your face model is usually within about one year of the truth but occasionally out by three, and you must keep out anyone under eighteen. If you set the cut-off at eighteen, every seventeen-year-old the model overestimates by a year gets in, which is exactly the failure the law is trying to prevent.

So you set the cut-off at twenty-five instead. If the model says twenty-five or more, the person goes straight through. If it says less, you do not refuse them, you send them to a stronger check, usually a document. A genuine seventeen-year-old would now have to be estimated eight years too old to slip past, which happens far less often than being estimated one year too old.

That gap between the legal age and the threshold is the buffer, and the threshold itself has a name in the regulator’s vocabulary: the challenge age. It is precisely the bouncer’s instinct, written as a rule. If you plainly look old enough, no document. If you might be young, document.

The problem you cannot wave away#

Now the hard part.

The nightclub bouncer sees your passport and forgets you. A website cannot forget, because it is a computer, and computers keep things unless somebody deliberately makes them not. A site that has seen your passport holds, in one place, the fact that a named person visited it. For most sites that is merely unpleasant; for the sites these laws are aimed at it is dangerous, because a list of names attached to pornography, or to a support forum for a stigmatized condition, is a blackmail file, and databases leak.

So the design problem is this: let the site be confident you are an adult, while making sure the site never learns your name and whoever checked your name never learns which site you were visiting.

That is achievable, and the pattern is called double blind. You get your age checked by one company, which knows who you are but not where you are going. It gives you a sealed note saying only “this person is over 18”. You hand the note to the website, which can tell the note is genuine because it recognizes the checker’s seal, cannot open it to find your name, and does not tell the checker you arrived. Two parties, each missing one half of the story, neither able to reconstruct the whole.

One Saturday morning#

Let us fix a person and keep her for the rest of the chapter.

Rosa Iyer was born on 12 September 2008. On Saturday 1 August 2026, the morning we care about, she is seventeen years and ten months old. She lives in Manchester. She is, in the language of the Online Safety Act, a child, because that Act defines a child as a person under the age of 18, with no gradations.

Rosa opens a pornography site on her phone. Because the site allows users in the United Kingdom, and the duty has been in force since January 2025, she sees no content. She sees a wall, offering several ways to prove her age, one of them a face scan, which needs no document at all.

She takes the selfie. The model returns an estimate of 20.3 years. The site’s challenge age is 25. Twenty point three is below twenty-five, so she is not let through; she is offered the document check instead. Rosa does not have a passport. She closes the tab.

Her brother Vikram Iyer was born on 3 June 2007, which makes him nineteen years and two months old that same morning. He is legally entitled to be there. He takes the selfie. The model returns 23.1. Also below twenty-five. He is also sent to the document check, and because he does have a passport, he spends forty seconds photographing it and then gets in.

Both outcomes are the system working as designed. The child was stopped, which is the point. The adult was made to show his passport to a stranger to look at pornography, which is not the point but is the price. We will put real numbers on how often that happens, and to whom, in the technical half.

Where the plain version stops being true#

“Verified” does not mean certain#

The word verification suggests a fact established beyond doubt. It is not.

A document check has at least four failure points. The document may be forged well enough to pass automated inspection. It may be genuine but belong to somebody else, most commonly an older sibling, which the face-match step is supposed to catch. The face match can itself be defeated by a presentation attack, holding a photograph, screen or mask up to the camera, or an injection attack, bypassing the camera and feeding a synthetic video straight into the app. And the underlying birth record can simply be wrong, which matters enormously in the large parts of the world where births are registered late or not at all.

The honest version: age verification produces a high-confidence assertion about an age, conditional on the document being genuine, the document belonging to the presenter, the capture being live, and the birth record being correct. Each of those conditions has a measurable failure rate, and none of them is zero.

The face does not know its own birthday#

A face model has never seen a birth certificate. It has seen photographs labelled with ages, and learned the statistical relationship between how skin, bone structure and expression appear and how many years the person has lived. That relationship is real but loose, and looser for some people than others. Two consequences follow, and marketing ignores both.

The first is that accuracy is not one number. A model excellent at telling a nine-year-old from a fifteen-year-old may be mediocre at telling a seventeen-year-old from a nineteen-year-old, which is the distinction the law cares about, and error grows with age: algorithms within about one and a half years for children are out by three or four for people in their sixties and seventies.

The second is that error has a direction, unevenly distributed. In the National Institute of Standards and Technology’s evaluation, teenagers tend to be estimated older than they are and older adults younger, and estimates vary systematically by demographic group. So the buffer that protects children in one population may be too small in another, and the friction imposed on adults is heavier for some groups than others.

Blocking is not the same as stopping#

An age check is a wall around one property, not a fence around a field.

Every check the law imposes applies only to services within its reach. A site incorporated abroad and ignoring the letters is not within reach in any practical sense until the regulator gets to it, and regulators get to things one at a time. Meanwhile a virtual private network, a paid service that makes your traffic appear to come from another country, costs a few pounds a month and takes two minutes to install. Children know this.

So the effect is not that children can no longer see this. It is that they can no longer see it on the large, compliant, advertising-funded services, and the traffic redistributes: some towards a virtual private network and the same site, some towards smaller sites nobody has caught. Whether the net effect is positive is an empirical question, and we will look at the data rather than assert an answer.

The buffer protects children by inconveniencing adults#

The buffer story in the plain version is true and incomplete. It is not a free win but a dial, and turning it towards protecting children always turns it towards inconveniencing adults. Raise the challenge age from twenty-five to twenty-eight and fewer under-eighteens slip through, while a great many more genuine twenty-five-year-olds face a document check they should not need. Lower it to twenty-one and adults sail through, and so do sixteen-year-olds who look old for their age. There is no threshold at which both error types vanish, because the underlying distributions overlap, and anyone claiming to have solved this has changed the subject.

“Nobody stores anything” is a design goal, not a fact#

Vendors say the selfie is deleted immediately, and often it genuinely is. But deleting the image is not the same as retaining nothing. A provider still typically keeps an audit log, because the regulator requires the service to demonstrate what it did, and may keep a hash of the face template to detect repeat attempts. The relying party keeps the result and the timestamp. A payment processor, if a card was used, keeps a merchant record. And the site’s server logs record an internet address and a time. None of that is the photograph, and all of it is data about a person visiting a site.

The honest version: privacy in age assurance is not achieved by promising deletion. It is achieved by architecture that structurally prevents one party from holding both halves of the sensitive pair, which are your identity and your destination. Deletion promises are a policy; architecture is a property. Chapter 55 covers the legal machinery that turns such promises into obligations.

The law is not one law#

It is tempting to think of “age verification law” as one thing that countries are adopting at different speeds. It is not. The United Kingdom regulates the outcome and lets you choose the method. American states mostly enumerate permitted methods in the statute. The European Union has largely not legislated a specific duty, working through a general platform obligation plus guidance plus a reference implementation. Australia has banned a category of account for a category of person rather than gating content. These are different legal shapes, not different speeds along one road, and a design that satisfies one may fail another.

The technical version#

The exact vocabulary, and where it is written down#

Start with the statute, because the United Kingdom is the one place where two of these terms have a legislative definition rather than an industry convention.

The Online Safety Act 2023, chapter 50, received Royal Assent on 26 October 2023. Section 230, headed “Age verification” and “Age estimation”, is short enough to give in substance. 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. And subsection (4) does something important: a measure requiring a user to self-declare their age, without more, is not to be regarded as either. The tick box is not merely weak in British law, it is definitionally outside the category.

Section 230 does not define “age assurance”. That word runs through Ofcom’s guidance and the industry as the umbrella covering both, but it is a regulatory and industry term rather than a statutory one: a convention, not a standard, in British law.

For the standard, go to ISO/IEC 27566-1:2025, “Information security, cybersecurity and privacy protection, Age assurance systems, Part 1: Framework”, published on 12 December 2025 by ISO/IEC JTC 1/SC 27, first edition, twenty-nine pages, and the first international standard on the subject. It establishes a framework and describes core characteristics, which published summaries of the standard group as functionality, performance, privacy, security and acceptability [UNVERIFIED: the five core characteristic headings of ISO/IEC 27566-1:2025, taken from published summaries rather than from the standard text]; it splits methods into verification, estimation and inference, the same three-way division we used in the plain half; and it promotes successive validation, the formal name for the layered step-up.

The table fixes the vocabulary, which press coverage uses loosely and specifications use precisely.

One more term, because Ofcom made it central in 2026. Age inference as practised by large social platforms is narrower and weaker than the ISO sense: analysing a user’s behaviour on the platform to guess whether they are a child. That is not inferring adulthood from a credit card; it is a behavioural model with no external evidence in it, and the British regulator has now said it does not accept it.

The method catalogue, and what each one actually checks#

Ofcom’s guidance sets out a non-exhaustive list of kinds of age assurance capable of being highly effective, and a shorter list that are not. These lists are the clearest published statement anywhere of what a regulator will and will not accept.

Capable of being highly effective: open banking, photo-identification matching, facial age estimation, mobile-network operator age checks, credit card checks, digital identity services and email-based age estimation. Not capable: self-declaration of age, age verification through online payment methods which do not require the user to be over 18 such as debit cards, and general contractual restrictions on use of the service by children.

Map those onto the five families and the mechanics become clear.

Document methods are photo-identification matching. The user captures a passport, driving licence or national identity card; the system authenticates it by checking printed security features, machine-readable zone check digits and, where the document has a chip, the digital signature over its data groups; then it captures a live selfie and performs a one-to-one biometric comparison against the portrait. The liveness step stops a photograph of a photograph. The date of birth comes from the document data, not the face.

Credit reference and database methods cover open banking, credit card checks, mobile-network operator checks and digital identity services. Open banking means the user authorizes their bank to confirm a narrow fact through a regulated interface. A credit card check works because issuers do not, as a rule, give credit to minors, which is why Ofcom accepts credit cards and rejects debit cards: a thirteen-year-old can hold a debit card. Mobile-network checks work in the United Kingdom because adult filters are on by default until the account holder proves their age to the network, so the network can answer that a line is unrestricted without disclosing anything else. Digital identity services means the certified providers described in chapter 52, reused here.

Facial age estimation takes a selfie, runs it through a model trained on faces of known age, and returns a number; the next section measures how good those numbers are.

Vouching is not on Ofcom’s list. The absence is not a prohibition, since the list is expressly non-exhaustive, but a service relying on vouching alone for an over-eighteen decision would be taking a serious risk. It appears in a better-supported role elsewhere: Australia’s Age Assurance Technology Trial, run by the Age Check Certification Scheme and published in final form on 31 August 2025, treated parental control and parental consent as distinct parts of its report alongside verification, estimation, inference and successive validation.

Hard identifiers means a government-issued number resolvable against a state population record: the simplest option technically and the most privacy-destructive, because the number is a universal correlation key across every other system using it. Where a jurisdiction mandates one, age assurance becomes a database lookup and double blindness becomes very hard, since the checker necessarily learns the exact individual. Chapter 50 covers what happens when a country builds identity around such a number.

Measuring facial age estimation properly#

The National Institute of Standards and Technology runs the only large independent evaluation, the Face Analysis Technology Evaluation, Age Estimation and Verification track, documented in NIST Interagency Report 8525 and updated as vendors submit algorithms. As of the August 2026 update it tests against around eleven million photographs from four operational sources: immigration applications, visas, border crossings and United States arrest mugshots.

It reports several metrics, and you need all of them. Mean absolute error, MAE, is the average size of the error in years ignoring direction, and is the number vendors quote. Mean error, ME, is the average signed error, telling you whether the algorithm systematically ages people up or down; two algorithms with identical MAE behave completely differently at a legal threshold if one consistently overestimates teenagers. Accuracy within three years, Acc(3), is the proportion of estimates within three years of the truth. Challenge-T evaluates the buffer directly: for a threshold T, someone estimated at or above T is accepted outright and everyone below is sent to a further check, and NIST notes that T equal to 25 is common where the legal age is 18, with a higher T needed for equivalent behaviour at a legal age of 21.

The numbers that matter are these. Across leading algorithms in the 2026 results, MAE is roughly 1.25 to 1.55 years for subjects aged 0 to 17, rising towards three to four years or more for older subjects. Yoti, one of the vendors, publishes its own white paper; the July 2025 edition reports a mean absolute error of 1.1 years for 13 to 17 year olds, 1.3 years for 6 to 12 year olds and 2.4 years across 6 to 70 year olds, with a true positive rate of 99.3 per cent for 13 to 17 year olds correctly estimated as under 21. Treat vendor figures and NIST figures as different kinds of evidence: the vendor tests on data it chose, NIST tests on data the vendor has never seen.

Now the part usually left out. NIST’s demographic breakdown for 2026 shows that at age 24 with a challenge age of 25, average acceptance across algorithms on application-quality images runs from about 51 per cent for Eastern European females to about 70 per cent for Eastern African males, and on border-quality images from about 47 to about 80 per cent. In terms of a real user: whether a twenty-four-year-old is waved through or sent to a passport check depends substantially on their demographic group and the camera they happen to be using. That is why Ofcom made fairness one of its four criteria rather than leaving it to good intentions.

The buffer arithmetic, worked all the way through#

Return to Rosa and Vikram on 1 August 2026 and put the published numbers on their morning.

Rosa is 17.89 years old. Her estimate came back 20.3, an overestimate of 2.41 years, well inside the normal range for a teenager, since NIST’s mean error data shows a broad tendency for younger subjects to be estimated older. Against a challenge age of 25 she is stepped up; against a challenge age of 18 she would have been admitted outright, which is the argument for the buffer in one line. Vikram is 19.16 years old, estimated 23.1, an overestimate of 3.94 years. He is also stepped up, and he is an adult entitled to walk in.

Now widen from these two to the population, using the NIST 2026 challenge threshold figures. For a genuine twenty-year-old, the proportion who clear the threshold outright, and so avoid the document check, is 12.0 per cent with Regula’s submitted algorithm at T equal to 25, 14.0 per cent with ROC’s and 16.4 per cent with Yoti’s. Raise the threshold to 28 and those figures fall to 3.1, 3.6 and 3.7 per cent respectively.

Read from the operator’s side, that means the following. With a challenge age of 25, between about 84 and 88 per cent of your twenty-year-old adult users will be sent to a document check. Raise it to 28 to squeeze out the last of the risk and between about 96 and 97 per cent of them will be. You have bought a small reduction in the number of teenagers slipping through by roughly tripling the number of young adults who must show a passport.

Legal age            18
Challenge age T      25          28
--------------------------------------------------
Age-20 subject clears outright:
  Regula-000       12.0%        3.1%
  ROC-002          14.0%        3.6%
  Yoti-004         16.4%        3.7%
--------------------------------------------------
So age-20 users sent to document step-up:
  Regula-000       88.0%       96.9%
  ROC-002          86.0%       96.4%
  Yoti-004         83.6%       96.3%
Source: NIST FATE AEV, figures as reported
        August 2026.

Two warnings about that block. “Clears outright” is measured against the challenge threshold, not the legal age: a twenty-year-old who fails a challenge age of twenty-five has not been judged a child, only judged not obviously over twenty-five. And these are per-algorithm figures from one evaluation at one moment, which move as vendors submit new models, so any deployment decision should be taken against the current report.

One more cost never appears in accuracy tables. Rosa had no passport and abandoned, and so does every adult without one. The document route is not universally available, and the population without a passport is not randomly distributed. An age assurance design that can only be completed by passport holders has quietly built an eligibility test out of document ownership.

The United Kingdom: the Act, the criteria and the dates#

Britain tried this once before and failed. Part 3 of the Digital Economy Act 2017 required commercial pornography sites to use robust age verification, named the British Board of Film Classification as regulator, and was delayed repeatedly over what counted as robust, over privacy objections to the dominant provider, and over a failure to notify the scheme to the European Commission. On 16 October 2019 the government abandoned it, and the Online Safety Act 2023 later repealed the provisions. That failure is why the second attempt is written as an outcome duty on services rather than a scheme run by a regulator.

The regime that replaced it is the most developed in the world as of August 2026, and worth going through precisely because everybody else is copying pieces of it. Two duties matter, and they sit in different Parts of the Online Safety Act 2023.

Section 81 applies to Part 5 services, meaning services that publish or display their own pornographic content rather than hosting other people’s. Section 81(2) creates a duty to ensure, by the use of age verification or age estimation or both, that children are not normally able to encounter that content. Section 81(3) is the standard: 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(4) and 81(5) add a record-keeping duty and a duty to publish a summary of that record, including which kinds of age check are used and how.

Section 12 applies to Part 3 user-to-user services likely to be accessed by children. Section 12(3)(a) requires proportionate systems to prevent children of any age encountering primary priority content that is harmful to children. Section 12(4) requires the provider to use age verification or age estimation to achieve that, and section 12(6) applies the same “highly effective at correctly determining whether or not a particular user is a child” standard. Section 12(5) contains the escape hatch that a great deal of the industry now relies on: the requirement does not apply if the terms of service prohibit that kind of content entirely and the policy applies to all users. That is why so many platforms responded by banning pornography outright rather than building an age gate.

Neither section says how. “Highly effective” is an outcome standard, and Ofcom filled in the meaning in its Statement on Age Assurance and Children’s Access of 16 January 2025, with separate guidance for Part 5 and Part 3 services. It sets four criteria, all of which a process must fulfil:

1. Technical accuracy: the degree to which the method can correctly determine the age of a user under test lab conditions. Ofcom’s practical steps include evaluating the method against appropriate metrics and, where age estimation is used, using a challenge age approach. That is the regulator mandating the buffer.
2. Robustness: the degree to which the method works in actual deployment contexts, tested in multiple environments, with steps taken to mitigate circumvention methods that are easily accessible to children.
3. Reliability: the degree to which the output is reproducible and derived from trustworthy evidence, with specific obligations where machine learning is involved, including monitoring for reproducibility after deployment and root-cause analysis when results go strange.
4. Fairness: the extent to which the method avoids or minimises bias and discriminatory outcomes, including training on data sets reflecting the diversity of the target population and evaluating against error parity.

Alongside them sit two principles rather than criteria: accessibility, meaning the method should work for all users regardless of characteristics, and interoperability, meaning systems should communicate using common and standardized formats.

One thing is conspicuously absent. Ofcom set no numerical thresholds; many consultation responses asked and did not get them. A service cannot demonstrate compliance by hitting a published false-acceptance rate, but must argue that its process as a whole meets four qualitative criteria. Practitioners find this genuinely difficult, and it is a fair criticism of the regime.

Each of the dates below changed something concrete.

Section 81 was commenced on 17 January 2025 by the Online Safety Act 2023 (Commencement No. 4) Regulations 2024, S.I. 2024/1333, regulation 2(1)(a), with no transition period for services already publishing pornographic content to United Kingdom users. The last date in the table is the one the public noticed, because 25 July 2025 is the day the large social platforms and the remaining pornography sites all put walls up at once.

What actually happened, measured#

Section 157 requires Ofcom to report on the use of age assurance. It did so on 15 July 2026, and this is the best evidence anywhere about what these laws do in practice. The findings support neither of the loud positions.

Volume first. Across a sample of 32 services operating in the United Kingdom, over 69 million age checks were completed between July and December 2025, a twenty-three-fold increase on the previous six months. Coverage rose too: 64 of the top 100 most popular pornography services had deployed age assurance as of June 2026, a further 10 were geo-blocking United Kingdom users entirely, and all of the top 10 had checks.

Effect on children next. The proportion of children who, on being asked to prove their age, met a highly effective check rose from 25 per cent in July 2025 to 43 per cent in January 2026. In Ofcom’s passive measurement study of 8 to 14 year olds, running from October 2025 to March 2026, 8 per cent of the panel visited a pornography service in a month, half of them only reaching sites with checks, and 4 per cent visiting at least one service with no checks at all. The durations are striking: 65 per cent of visits lasted under 10 seconds and 87 per cent under 30, which Ofcom reads as children bouncing between sites looking for a way in rather than viewing content.

Ofcom then listed four failures. First, too many pornography sites still have no checks. Second, children find them through search: 62 per cent of the children who reached a pornography service had used a search engine immediately beforehand at least once, and Ofcom found 33 per cent of first-page Google results and 54 per cent of Bing results pointed to sites without checks or other protections. Third, the age inference models used by some large social platforms are, in Ofcom’s words, subject to serious doubts; it opened a formal investigation into TikTok’s compliance with section 12 on 16 July 2026 and told the industry that age inference will not support an under-sixteen ban. Fourth, over one in ten 15 to 17 year olds were still visiting and spending time on three of the most used dating apps in December 2025 despite those apps having checks.

Enforcement is real. By the July 2026 report Ofcom had opened 23 investigations into the providers of 88 adult services, and 73 per cent of those services had either implemented age assurance or blocked United Kingdom users as a result. Seven providers covering 24 sites had been fined.

Cost, finally, because the industry predicted ruin. Ofcom’s April 2025 statement estimated per-check costs of 5 to 30 pence; the July 2026 report found actual costs towards the lower end, with a median across analysed providers of around 6 pence. Age assurance turned out to be cheap. The expensive parts were user abandonment at the wall and the engineering to make the wall work at all.

Where Britain is heading: the government announced on 15 June 2026 that it will ban social media for under-sixteens on the Australian model, with rules expected in 2027; Ofcom is to give Parliament a rapid assessment by the end of October 2026 of what a highly effective over-sixteen check looks like, and a statutory report on app store protections by January 2027. All of that is announced rather than in force as of August 2026.

The United States: Paxton, and the split that remains#

American age verification law began in Louisiana. Act 440 of 2022, effective 1 January 2023, required age verification on sites where a substantial portion, defined as 33.33 per cent, of the content is material harmful to minors. The drafting was widely copied: the one-third threshold and the phrase “material harmful to minors” recur across most later state statutes.

Texas enacted House Bill 1181 in 2023, codified at Texas Civil Practice and Remedies Code section 129B.001 and following. It applies to a commercial entity that knowingly and intentionally publishes or distributes material on an internet website more than one-third of which is sexual material harmful to minors, and requires reasonable age verification that the visitor is 18 or older. The Free Speech Coalition, the pornography industry’s trade body, challenged it; the Fifth Circuit held that no heightened scrutiny applied at all; the Supreme Court granted certiorari, heard argument on 15 January 2025 and decided on 27 June 2025.

Free Speech Coalition, Inc. v Paxton, No. 23-1122, 606 U.S. ____ (2025), held that H.B. 1181 triggers, and survives, review under intermediate scrutiny because it only incidentally burdens the protected speech of adults. Justice Thomas delivered the opinion, joined by Chief Justice Roberts and Justices Alito, Gorsuch, Kavanaugh and Barrett. Justice Kagan dissented, joined by Justices Sotomayor and Jackson.

The reasoning is widely misdescribed, so state it accurately. The Court did not hold that pornography is unprotected. It held that content obscene to minors but not to adults is, in its phrase, protected in part and unprotected in part: states may prevent minors from accessing it, following Ginsberg v New York, 390 U.S. 629 (1968), but may not prevent adults from doing so, following Butler v Michigan, 352 U.S. 380 (1957). The power to prevent minors from accessing such material therefore includes the power to require proof of age, verification being the ordinary and appropriate means of enforcing an age-based limit, as it already is for alcohol, firearms and driving licences. So no person, adult or child, has a First Amendment right to access such speech without first submitting proof of age; the burden on adults is incidental; and incidental burdens get intermediate scrutiny under Turner Broadcasting System, Inc. v FCC, 512 U.S. 622 (1994). Applying that, the Court found the law advances an important interest without burdening substantially more speech than necessary, the permitted methods of government identification and transactional data being established and already in use. It also rejected Texas’s argument for rational-basis review, which would ignore the real burden on adults, so the decision sits deliberately in the middle.

Two things did not happen on 27 June 2025, and both are constantly reported as if they had. The Court did not bless age verification generally; its holding is about material obscene as to minors, a narrow and long-established category. And it did not resolve social media age verification, which involves speech fully protected for minors as well as adults, and where lower courts have continued to enjoin state laws on First Amendment grounds.

That is the first split. The second is architectural: American states have divided over where in the stack the check should sit.

The site-level approach puts the duty on the website showing the content. As of August 2026 the Free Speech Coalition’s list of state age verification laws for adult content showed 27 states with effective dates, from Louisiana on 1 January 2023 to West Virginia on 12 June 2026 and Iowa on 1 July 2026; the same page’s summary line still said twenty-five, so treat the count as a moving number.

The device-level approach puts the duty on the app store or the operating system, which knows the account holder and can hand every app a signal instead of every app checking separately. Texas Senate Bill 2420, the App Store Accountability Act, is the leading example. It requires app stores to verify a user’s age category and obtain parental consent before a minor downloads an app or makes an in-app purchase, with four categories: under 13, 13 to 15, 16 to 17, and 18 and over. Scheduled for 1 January 2026, it was preliminarily enjoined in December 2025 by Judge Robert Pitman in the Western District of Texas on First Amendment grounds, then took effect after the injunction was lifted on appeal, with the Supreme Court declining to block it in late May 2026. Utah’s equivalent applies to app stores and developers from 6 May 2026 with enforcement from 31 December 2026; Louisiana’s takes effect on 1 July 2026 and California’s on 1 January 2027. North Dakota went further with Senate Bill 2380, effective 1 August 2026, requiring device manufacturers, operating systems and app stores to estimate the age of the primary user and transmit a digital age signal that websites and apps must act on.

The two designs pull in opposite directions on privacy. Site-level checks mean every site sees evidence about you. Device-level checks mean one company, the platform holding your account, sees everything, and every app learns a fact it did not have before. Neither is obviously better.

The European Union: an old directive, a general duty, and a reference build#

Europe arrived from an unusual direction: it has had a relevant obligation since 2018 and has only recently worked out what to do with it.

Directive (EU) 2018/1808 of 14 November 2018 amended the Audiovisual Media Services Directive, Directive 2010/13/EU, inserting Article 6a. That article requires member states to take appropriate measures to ensure that audiovisual media services which may impair the physical, mental or moral development of minors are only made available so that minors will not normally hear or see them, names age verification tools among the permitted measures, and says the most harmful content, such as gratuitous violence and pornography, shall be subject to the strictest measures. Article 6a(2) adds a routinely ignored rule: personal data of minors collected under paragraph 1 shall not be processed for commercial purposes such as direct marketing, profiling and behaviourally targeted advertising. Article 28b imposes parallel obligations on video-sharing platform services, and Article 28b(3)(f) lists operating age verification systems among them.

The Digital Services Act then added a general obligation. Article 28(1) requires providers of online platforms accessible to minors to put in place appropriate and proportionate measures to ensure a high level of privacy, safety and security of minors on their service, and does not mention age verification at all. That is the crucial textual point: the European duty is an outcome duty with no named method, and everything specific has been built on top of it by guidance rather than by legislation.

On 14 July 2025 the Commission did two things at once. It published guidelines on the protection of minors under the DSA, applying to all online platforms accessible to minors except micro and small enterprises. They recommend age assurance methods that are accurate, reliable, robust, non-intrusive and non-discriminatory; age verification specifically for adult content such as pornography and gambling, or where national rules set a minimum age; and age estimation elsewhere, such as where terms and conditions set a minimum age below 18. Following them is voluntary and guarantees nothing, but the Commission will use them to assess compliance with Article 28(1).

The same day it released the first version of an EU white-label age-verification blueprint, an open-source reference implementation aligned with the European Digital Identity Wallet architecture, which chapter 51 describes. The design is explicitly double blind and single use: the issuer verifies the user’s age using detailed personal data such as date of birth, the online service receives only a proof that the user is over 18, the proof provider is not told which service the proof is used at, and each proof is used once to prevent cross-service correlation.

On 29 April 2026 the Commission adopted a non-binding Recommendation establishing a common EU-wide framework for age verification technologies. Its headline is a deadline: at least one compliant solution in each member state by 31 December 2026. It states that methods should include technical safeguards against privacy risks such as tracking of online activity, naming zero-knowledge proofs. And it draws a line under national social media bans: a member state considering technical measures restricting minors’ access must notify the Commission before adoption, and failure to notify is a procedural defect rendering the measure unenforceable against individuals in national proceedings. As of August 2026 France, Spain and Germany were each considering minimum ages of their own.

France went its own way and got there first on the privacy question. The CNIL, the French data protection authority, published a recommendation on 22 September 2022, “Online age verification: balancing privacy and the protection of minors”, setting out double anonymity: the site must not learn the user’s identity, and the age verification provider must not learn which site the user is visiting. Arcom, the audiovisual regulator, turned that into a binding technical standard published in October 2024 under the SREN law, mandatory from 11 April 2025. Under it, relevant sites must not directly collect identity or date of birth data at all, must offer at least one double-anonymity solution, must offer at least two different double-anonymity methods for generating a proof, for example one based on identity documents and one on age estimation, and the double-anonymity system must be available to at least 80 per cent of the adult population resident in France. That last requirement is a coverage floor, and it is the only one of its kind in any regime described here.

Privacy-preserving designs: double blind, tokenised, zero-knowledge#

Now the engineering that makes the conflict survivable: three distinct techniques, frequently confused, solving different problems.

Double blind is an architecture, not a cryptographic primitive. It splits knowledge so that neither party holds both halves of the dangerous pair: the provider learns who you are and not where you are going, the relying party learns you are over eighteen and not who you are. It is enforced by never sending the relying party’s identity to the provider, and never your identity to the relying party.

  You                Provider              Website
   |                    |                     |
   |-- 1 prove age ---->|                     |
   |    (passport,      |                     |
   |     or face)       |                     |
   |                    |                     |
   |<-- 2 signed -------|                     |
   |    proof: "over18" |                     |
   |    no name in it   |                     |
   |                    |                     |
   |----------- 3 present proof ------------->|
   |                    |                     |
   |                    |   4 check signature |
   |                    |     against known   |
   |                    |     issuer key      |
   |<---------- 5 access granted -------------|

  Provider never learns step 3 happened.
  Website never learns your name.

The weak points are where you would expect. If the proof is fetched at the moment of use, the provider sees the timing and can often infer the destination, which is why the European blueprint insists on proofs issued in advance. If the same proof goes to two sites, they can collude to link you, which is why single use matters. If provider and relying party are the same corporate group, the architecture is decorative.

Tokenised age assurance turns the result of a check into a reusable credential held by the user, so the check is not repeated at every site. euCONSENT ASBL, a non-profit based in Brussels, built AgeAware for exactly this: an interoperability network where a user who completes a check with one participating provider accepts a signed token confirming their age status, which other participants accept as evidence. It runs on a double-blind chain of trust with every transaction signed and auditable, and went live in November 2025. Commercial equivalents exist: Yoti’s Age Connect binds an anonymous age token to a device using passkeys, and k-ID’s OpenAge Initiative uses an AgeKey with the same double-blind property, so the service cannot see which provider did the original check and the provider cannot see where the key is reused.

Reuse is the largest usability win available. If one check a year serves a hundred sites, the friction argument mostly evaporates. It is also where correlation risk concentrates, because a reused token can be tracked unless it is deliberately made unlinkable.

Zero-knowledge proofs make unlinkability a property rather than a promise. You prove a statement without revealing anything beyond its truth: that the date of birth inside a credential signed by a government issuer implies an age over eighteen, without revealing the date of birth, the document number or anything else, and without producing a value recognizable on a second use.

This stopped being theoretical in 2025. On 3 July 2025 Google open-sourced its zero-knowledge proof libraries as longfellow-zk, built to produce proofs about existing credential formats, including ISO mobile driving licences, JSON Web Tokens and W3C Verifiable Credentials, which chapter 53 describes, rather than requiring a new credential type. That is what made it useful: it works over the documents that already exist. A relying party asks for a credential of format mso_mdoc_zk and names the zero-knowledge system, the circuit hash and the number of attributes it needs.

{
  "id": "cred1",
  "format": "mso_mdoc_zk",
  "meta": {
    "doctype_value": "org.iso.18013.5.1.mDL",
    "zk_system_type": [
      {
        "system": "longfellow-libzk-v1",
        "num_attributes": 1,
        "version": 6
      }
    ],
    "verifier_message": "challenge"
  }
}

The relying party validates the proof against the issuer’s certificate authority certificates and learns one bit. It does not learn the date of birth, and cannot recognize the same person returning tomorrow. The circuit hash pins the exact statement proved, which is what stops a verifier quietly asking for more than it claimed.

Be precise about the limits. This removes correlation risk in the credential. It does not remove network-level correlation, because your internet address, browser fingerprint and login cookie are all still there, and it does not stop a relying party asking for a different, more revealing credential instead. Cryptography closes one channel thoroughly and leaves the rest of the system as leaky as it was.

The layering is the point. A serious 2026 design uses all four: a double-blind split, so no party sees identity and destination together; a reusable token, so the strong check happens once; a zero-knowledge proof, so the token discloses one bit; and single use, so two sites cannot compare notes.

Circumvention, and what the data actually shows#

Here the argument gets loud and the evidence gets thin, so we separate measured claims from inferred ones.

Measured, and not seriously disputed. On 25 July 2025, the day the Protection of Children Codes came into force, virtual private network sign-ups spiked violently. Proton VPN reported a surge of over 1,400 per cent in United Kingdom sign-ups within hours, half of the top ten free apps in the United Kingdom Apple App Store that morning were virtual private network apps, and Proton said publicly that the surge was sustained and much higher than when France lost access to adult content.

Measured, and more useful, because it covers the following year rather than the following day. Ofcom’s July 2026 report estimates daily United Kingdom virtual private network users at 2.2 million against 1.2 million before 25 July 2025, close to a doubling and sustained rather than a spike. It also reports a fall of roughly a third in estimated daily visitors to pornography services over the same period, and that 5 per cent of children in its study said they had used a virtual private network, for a variety of activities rather than principally for adult content.

Inferred, and contested. What proportion of the traffic fall is children actually stopped, as against adults who gave up at the wall or adults who moved behind a virtual private network. Nobody can measure this directly, because the point of a virtual private network is that the traffic no longer looks British. The Age Verification Providers Association argued in September 2025 that most of the fall reflects adults deterred or switching rather than children blocked. Internet Matters, a children’s online safety organization, found no rise in children’s virtual private network use after the checks began, and that only 16 per cent of children who had used one used it to reach restricted adult content. Those findings are not strictly contradictory, since one is about adults and the other about children, but they are cited as if they were, and you should be suspicious of anybody quoting one without the other.

Measured, and the most interesting finding of all. Ofcom’s passive study found visits by 8 to 14 year olds to pornography services were overwhelmingly very short, 65 per cent under ten seconds and 87 per cent under thirty, with panellists visiting on average five different services in a month and 257 different services across the panel. The natural reading, and Ofcom’s, is that children are searching for a way in rather than finding one: a partial success and a partial indictment, because the walls work on the sites that have them and the long tail is unwalled.

Three lessons follow. Displacement is real and should be planned for, because a rule covering the largest hundred services moves demand to service one hundred and one, and Ofcom’s answer of enforcement against the long tail plus pressure on search engines is the only structurally sound one. Circumvention rates are neither a reason to do nothing nor evidence of success: these measures raise the effort required, and effort thresholds bite hard on casual access and lightly on determined access, as age controls always have. And the regulator noticed the second-order effect, with guidance that services should not host or permit content encouraging child users to circumvent age assurance, for example by linking to a virtual private network.

Building one: the decision procedure and the honest cost model#

Here is the shape of a defensible 2026 design, as a procedure rather than a product.

1. Determine your obligations, per market.
     UK:  Part 3 or Part 5?  s.12 or s.81?
     US:  which states, at what threshold?
     EU:  Art 28 platform?  AVMSD service?
2. Decide the decision you need, not the data.
     "over 18" is one bit.  Ask for one bit.
3. Order methods cheapest-first, then step up.
     a. reusable token, if presented
     b. facial age estimation, challenge age T
     c. document plus liveness
     d. human review, appeal route
4. Set T from current NIST figures, not folklore.
     legal age 18 -> T of 25 is the convention
     legal age 21 -> T must be higher
5. Never store what you did not need.
     keep: result, timestamp, method, audit id
     drop: image, document, date of birth
6. Publish the summary record.  s.81(5) requires
   it for Part 5 services in the UK.
7. Provide a route for people with no document
   and no bank account.  If you cannot, say so.

Step 3 is where the money is. Ofcom’s measured median per-check cost of about 6 pence, against an estimated 5 to 30 pence, tells you the direct cost of checking is not the constraint. Abandonment is: every user who reaches step 3c and leaves is a lost user, and that number dwarfs the pennies, which is why the reusable token in step 3a matters commercially as well as ethically.

Step 7 is the one most designs skip and the one a regulator will ask about, because Ofcom’s accessibility principle says the method should work for all users regardless of their characteristics. A flow whose only fallback is a passport has silently excluded everybody without one.

Two open questions to close on. Experts disagree about whether device-level or site-level checking is the better architecture: one camp says a single check at the operating system serves every app, spares thousands of small sites from building identity pipelines and is the only way to reach the long tail, while the other says it concentrates enormous power in two American platform companies, hands every app a tracking signal and makes the operating system a censor. Both concerns are legitimate, the United States is running the experiment, and the answer will be empirical.

Experts also disagree about whether behavioural age inference can be made highly effective. Platforms argue a model with years of behavioural data knows more about whether a user is thirteen than any single selfie does. Ofcom’s position as of July 2026 is that it does not accept this without reliable and compelling evidence, that such models appear to miss significant numbers of children, and that they will not support an under-sixteen ban. The disagreement is about evidence rather than principle, and the platforms hold the evidence.

54.98 Common wrong ideas#

Wrong: Age verification and age assurance are two names for the same thing. Right: Age verification means measuring exact age against a record, defined in section 230(2) of the Online Safety Act 2023; age estimation means estimating an age or range, section 230(3); age inference means deriving age from another verified fact; and age assurance is the umbrella covering all three, used throughout Ofcom guidance and ISO/IEC 27566-1:2025 but not defined in the Act.

Wrong: A tick box saying “I am over 18” is a weak form of age verification. Right: It is not a form of it at all in British law, because section 230(4) of the Online Safety Act 2023 provides that a measure requiring a user to self-declare their age, without more, is not to be regarded as age verification or age estimation, and Ofcom lists self-declaration among the methods not capable of being highly effective.

Wrong: Facial age estimation with a mean absolute error of about one year can enforce an eighteen limit directly. Right: A one-year average error still leaves a substantial tail of seventeen-year-olds estimated over eighteen, which is why Ofcom requires a challenge age approach where estimation is used, why 25 is conventional for a legal age of 18, and why NIST notes a legal age of 21 would need a higher threshold.

Wrong: The Supreme Court in Free Speech Coalition v Paxton upheld age verification for the internet. Right: The Court held on 27 June 2025, by six to three with Justice Thomas writing, that Texas H.B. 1181 triggers and survives intermediate scrutiny because it only incidentally burdens adults’ protected speech; the holding concerns material obscene as to minors but not adults, does not address fully protected speech such as social media, and rejected Texas’s argument for rational-basis review.

Wrong: The European Union has passed an age verification law. Right: Article 28(1) of the Digital Services Act imposes a general duty to ensure a high level of privacy, safety and security of minors and never mentions age verification; the specifics come from the Commission’s guidelines and blueprint of 14 July 2025 and the non-binding Recommendation of 29 April 2026 seeking one compliant solution per member state by 31 December 2026, alongside Article 6a and Article 28b of the Audiovisual Media Services Directive as inserted by Directive (EU) 2018/1808.

Wrong: Privacy-preserving age checks mean the provider deletes your selfie. Right: Deletion is a policy that can be reversed or breached; the structural protections are double blind architecture so no single party holds both your identity and your destination, single-use proofs so two sites cannot correlate you, tokenised reuse so the strong check happens once, and zero-knowledge proofs so the credential discloses one bit, as in Google’s longfellow-zk libraries of 3 July 2025.

Wrong: Age checks failed in the United Kingdom because everybody just used a virtual private network. Right: Ofcom’s report of 15 July 2026 found daily United Kingdom virtual private network users rose from about 1.2 million to about 2.2 million, a real and sustained rise, alongside over 69 million checks across 32 services in six months, children encountering highly effective checks rising from 25 to 43 per cent, and 64 of the top 100 pornography services checking by June 2026, so circumvention and effect are both substantial.

Wrong: Large platforms already know how old their users are from their behaviour, so behavioural age inference is enough. Right: Ofcom concluded in July 2026 that it has serious doubts about these models, that they are not in its guidance as capable of being highly effective, that some companies may be missing significant numbers of children, and that such methods will not support the under-sixteen restrictions announced on 15 June 2026; it opened a formal investigation into TikTok on 16 July 2026.

Wrong: Once a country requires age checks, children can no longer reach the content. Right: Ofcom’s passive measurement of 8 to 14 year olds found 8 per cent visited a pornography service in a month and 4 per cent reached at least one with no checks, 62 per cent had arrived straight after a search engine, and 33 per cent of first-page Google and 54 per cent of Bing results pointed to unchecked sites, so the effect is redistribution towards the unregulated long tail rather than elimination.

54.99 Chapter summary in 20 lines#

1. Age verification measures exact age against a record, age estimation guesses it from a characteristic, age inference derives it from another verified fact, and age assurance is the umbrella term.
2. Section 230 of the Online Safety Act 2023 defines age verification and age estimation, and section 230(4) excludes self-declaration from both, so a tick box is not a weak check in British law but none at all.
3. ISO/IEC 27566-1:2025, published 12 December 2025 as the first international standard on age assurance, frames systems around functionality, performance, privacy, security and acceptability, and promotes successive validation, a layered step-up.
4. The five method families are documents with liveness, records held by banks, card issuers and mobile networks, facial age estimation, vouching by a verified adult, and hard government identifiers.
5. Ofcom lists open banking, photo-ID matching, facial age estimation, mobile-network checks, credit card checks, digital identity services and email-based estimation as capable of being highly effective, and rejects self-declaration and debit card checks.
6. Facial age estimation accuracy must be read through mean absolute error, mean error, accuracy within three years and the challenge threshold together, because one headline figure hides the direction of error and its variation by age and group.
7. Leading algorithms in NIST’s 2026 evaluation show mean absolute error of about 1.25 to 1.55 years under age 18, rising to three or four years for older subjects; Yoti’s July 2025 white paper reports 1.1 years for 13 to 17 year olds.
8. The buffer, called a challenge age by Ofcom, sets the pass threshold well above the legal age, 25 being conventional for a legal age of 18, with everyone below it stepped up rather than refused.
9. Raising the challenge age from 25 to 28 cuts the proportion of genuine twenty-year-olds clearing outright from 12.0 to 16.4 per cent down to 3.1 to 3.7 per cent, so extra child protection is bought with much more adult friction.
10. Section 81 of the Online Safety Act 2023 requires Part 5 pornography publishers to use highly effective age verification or estimation, and was commenced on 17 January 2025 by S.I. 2024/1333.
11. Section 12 requires Part 3 user-to-user services likely to be accessed by children to do the same for primary priority content, unless their terms ban that content for all users, which is why many platforms banned rather than gated.
12. Ofcom’s guidance of 16 January 2025 sets four criteria, technical accuracy, robustness, reliability and fairness, plus the principles of accessibility and interoperability, and deliberately publishes no numerical thresholds.
13. The Protection of Children Codes were published on 24 April 2025 and came into force on 25 July 2025, the date the walls appeared across the British internet.
14. Ofcom’s report of 15 July 2026 found over 69 million age checks across 32 services between July and December 2025, a twenty-three-fold rise, with 64 of the top 100 pornography services checking by June 2026 and seven providers fined over four million pounds in total.
15. Free Speech Coalition, Inc. v Paxton, decided 27 June 2025 by six to three, held that Texas H.B. 1181 triggers and survives intermediate scrutiny because proof of age only incidentally burdens adults’ access to speech obscene as to minors.
16. American law is split between site-level statutes in about 27 states and device-level app store statutes in Texas, Utah, Louisiana, California and North Dakota, concentrating privacy risk in opposite places.
17. Article 6a of the Audiovisual Media Services Directive, inserted by Directive (EU) 2018/1808, has required protection of minors since 2018 and bars commercial processing of minors’ data collected for that purpose.
18. Article 28(1) of the Digital Services Act states an outcome duty with no named method, filled in by the Commission’s guidelines and blueprint of 14 July 2025 and its Recommendation of 29 April 2026 seeking a compliant solution in every member state by 31 December 2026.
19. Privacy-preserving age assurance layers four things: double-blind architecture, tokenised reuse as in euCONSENT’s AgeAware network live from November 2025, zero-knowledge proofs as in Google’s longfellow-zk libraries of 3 July 2025, and single-use proofs against cross-site correlation.
20. Ofcom measured daily United Kingdom virtual private network users rising from about 1.2 to about 2.2 million after 25 July 2025 and pornography visits falling by roughly a third, showing real effect and real circumvention, and settling neither side of the argument.

Chapter sources: Online Safety Act 2023, chapter 50, Royal Assent 26 October 2023, in particular section 12(3)(a), (4), (5) and (6) on children’s safety duties for Part 3 services, section 81(2) to (5) on regulated provider pornographic content, section 157 on the report on the use of age assurance, section 230 defining age verification and age estimation and excluding self-declaration, and section 236 defining child and pornographic content, with the Online Safety Act 2023 (Commencement No. 4) Regulations 2024, S.I. 2024/1333, regulation 2(1)(a) commencing section 81 on 17 January 2025; Ofcom, Statement: Age Assurance and Children’s Access, 16 January 2025, last updated 24 April 2025, with its Part 5 guidance of 16 January 2025 and its Part 3 Guidance on highly effective age assurance, for the four criteria, the accessibility and interoperability principles, the lists of methods capable and not capable of being highly effective, the challenge age instruction and the instruction not to host content encouraging circumvention; Ofcom’s Protection of Children Codes of Practice, published 24 April 2025 and in force 25 July 2025, with the children’s access assessment deadline of 16 April 2025 and risk assessment deadline of 24 July 2025; Ofcom, Report on the use of age assurance, prepared under section 157 and published 15 July 2026, with its news release of the same date, for the 69 million checks across 32 services, the twenty-three-fold rise, the increase from 25 to 43 per cent, the 64 of the top 100 services with checks and 10 geo-blocking as of June 2026, the Children’s Passive Online Measurement study of October 2025 to March 2026, the search-engine and first-page-results findings, the 23 investigations into 88 adult services with 73 per cent compliance and the seven named fines, the virtual private network user estimates of 1.2 and 2.2 million, the dating services finding, the TikTok investigation of 16 July 2026 and the per-check costs of 5 to 30 pence against an observed median near 6 pence; ISO/IEC 27566-1:2025, Age assurance systems, Part 1: Framework, edition 1, published 12 December 2025 by ISO/IEC JTC 1/SC 27, 29 pages; NIST Interagency Report 8525, Face Analysis Technology Evaluation: Age Estimation and Verification, and the FATE AEV results as reported in August 2026, for the corpus of about eleven million images, the metrics mean absolute error, mean error, Acc(3) and Challenge-T, the note that T of 25 is common for a legal age of 18, the error ranges by age band, the age-20 clearance rates at T of 25 and 28 for Regula-000, ROC-002 and Yoti-004, and the demographic acceptance spreads at age 24; the Yoti Facial Age Estimation white paper of July 2025; Free Speech Coalition, Inc. v Paxton, No. 23-1122, 606 U.S. ____ (2025), argued 15 January 2025, decided 27 June 2025, Thomas J. for the Court with Roberts C.J. and Alito, Gorsuch, Kavanaugh and Barrett JJ., Kagan J. dissenting with Sotomayor and Jackson JJ., affirming 95 F. 4th 263, and its treatment of Ginsberg v New York, 390 U.S. 629 (1968), Butler v Michigan, 352 U.S. 380 (1957) and Turner Broadcasting System, Inc. v FCC, 512 U.S. 622 (1994) and 520 U.S. 180 (1997), with Texas Civil Practice and Remedies Code section 129B.001 and following; Louisiana Act 440 of 2022, effective 1 January 2023; Texas Senate Bill 2420, enjoined in December 2025 before taking effect on appeal with the Supreme Court declining to block it in late May 2026, the Utah, Louisiana and California app store statutes and North Dakota Senate Bill 2380 effective 1 August 2026; the Free Speech Coalition state age verification law tracker as consulted in August 2026; Directive 2010/13/EU as amended by Directive (EU) 2018/1808 of 14 November 2018, Article 6a paragraphs 1 to 4 and Article 28b including paragraph 3(f); Regulation (EU) 2022/2065, Article 28(1); the European Commission guidelines on the protection of minors and the EU white-label age-verification blueprint, both 14 July 2025; the Commission Recommendation on a common EU-wide framework for age verification technologies of 29 April 2026; the CNIL recommendation of 22 September 2022 on double anonymity and the Arcom technical standard of October 2024 under the SREN law, mandatory from 11 April 2025, with its 80 per cent coverage floor; euCONSENT ASBL’s AgeAware network, live from November 2025, with Yoti Age Connect and the k-ID OpenAge Initiative; Google’s announcement of 3 July 2025 open-sourcing the longfellow-zk libraries and the Google Wallet documentation for the mso_mdoc_zk format; Digital Economy Act 2017 Part 3 with the British Board of Film Classification as designated regulator, abandoned on 16 October 2019 and later repealed by the Online Safety Act 2023; Australia’s Online Safety Amendment (Social Media Minimum Age) Act 2024, assented 10 December 2024 and in force 10 December 2025, with the Age Assurance Technology Trial final report of 31 August 2025; the United Kingdom government announcement of 15 June 2026; and Proton VPN’s reported sign-up surge of over 1,400 per cent on 25 July 2025 with the Age Verification Providers Association analysis of September 2025 and Internet Matters research on children’s virtual private network use.