From the journal

Age verification for AI apps and digital platforms: lawful checks, privacy and vendor responsibility

An AI service should choose its age check after deciding which users may use each feature, which child-data rules apply and what evidence the applicable law requires. A general assistant, a persistent companion, a public feed and a sexual-content generator can produce four different answers inside one product. The website and mobile app also need a consistent permission system, even when their available age signals differ.

30 min read

The practical aim is to establish the necessary age or authority with reliable evidence and limited data. That may mean a protected experience for all users, a signed age-range assertion, or an independently tested estimate with another route for uncertain cases. The supplier must demonstrate that its method meets the particular duty. The platform retains the responsibilities assigned to it by law, including when a store, operating system or specialist verifier supplies the signal. GDPR Article 28, Articles 14–15 of Brazil's Digital Statute of Children and Adolescents (ECA Digital) and Ofcom's vendor guidance illustrate that retained responsibility.

Start with five separate decisions

An age result should feed a permission record with five distinct questions:

  1. Access: may this person use this feature or receive this content?
  2. Data processing: what lawful basis applies, and does a child-consent rule require consent from a person with the necessary legal authority?
  3. Parental authority: is the person giving consent legally entitled to act for this child, and for this purpose?
  4. Contract: can this person enter the subscription or other obligation under the applicable capacity rules?
  5. Distribution: what does the store, operating system (OS) or app policy require for installation, purchases, updates and the app's actual content?

The age thresholds for these decisions differ. GDPR Article 8 addresses consent-based services offered directly to children; the national age for a child to consent independently is 15 in France and 16 in Ireland. Under the French rule, consent below 15 is given jointly by the child and the holder or holders of parental authority. UK GDPR uses 13 for its corresponding rule. Those provisions have defined scope and exceptions. They do not create a single account-access age or settle contractual capacity. Under the US Children's Online Privacy Protection Act (COPPA), covered collection from under-13s normally requires verifiable parental consent; a primarily child-directed service cannot escape that duty through a birthday field. GDPR Article 8; France Article 45; Ireland section 31; UK GDPR Article 8; COPPA sections 312.2–312.5.

A card, phone contract or OS account can belong to an adult while a child uses the app. Verification must therefore address the connection between the evidence, the relevant account and the actual user. Parental permission requires additional authority and purpose-specific consent evidence. Ofcom method guidance; COPPA consent methods.

Service features determine the age duty

Use a feature-and-territory assessment before writing a supplier specification. The following is a list of duties; method selection follows separately.

  • Private general assistance. Legal question: Does child-directed, actual-knowledge, likely-access or local child-data law apply? Consequence for the design: Identify the audience and a lawful data route. Kenya's statutory age/consent mechanisms and Brazil's likely-access framework can matter without a public feed. Kenya section 33; Brazil Articles 1–2.
  • Persistent companion memory, emotional prompts or relationship simulation. Legal question: Does the feature meet a companion definition, use a prohibited behaviour or track a child? Consequence for the design: Map each behaviour. New York's 2027 unsafe-feature rule, California's staged companion duties and India's 2027 monitoring restrictions use different tests. New York Article 48; California SB1119; India section 9.
  • Profiles, inter-user sharing, recommendations or feeds. Legal question: Is this a covered platform, social service, user-to-user service or addictive feature? Consequence for the design: Apply the actual definition, territorial connection, designation and exemptions. Australia, the EU, the United Arab Emirates (UAE), California and Mexico do not use interchangeable categories. DSA Articles 2–3,19,28; Australia rules; UAE Resolution 106; California AB1709; Mexico decree.
  • Restricted sexual content or generation. Legal question: Is the material prohibited, or available only with effective adult-access controls? Consequence for the design: Assess generated text, images and video separately. Brazil expressly reaches erotic AI dialogue; UK Part 5's provider-pornography route excludes text-only material. Other duties may still cover that text. Brazil decree Article 16; UK sections 79–81.
  • Mobile installation, purchase or material update. Legal question: Which store/OS and developer duties apply? Consequence for the design: Obtain legally permitted signals and permissions, preserve their provenance and support direct web access. A content rating concerns the app's content. California AB1856; Singapore ADS Code; Apple; Google.
  • New verifier, model, inference source or processing location. Legal question: Does the change affect method legality, data roles, biometrics, transfers or accuracy? Consequence for the design: Repeat the affected assessment before deployment and require evidence from the supplier. GDPR Articles 28,35,44; New Zealand biometric code; DIFC Regulation 10.

Territory also follows the particular instrument. GDPR can attach through EU establishment, offering services to people in the Union or monitoring them there. UK online-safety rules have their own UK-link tests. Brazil's child framework can reach foreign suppliers serving the covered Brazilian audience. South Africa's Protection of Personal Information Act (POPIA) uses domicile or specified means of processing in the Republic. Record the connection for each regime rather than use company nationality as a shortcut. GDPR Article 3; UK sections 4,80; Brazil Article 1; POPIA section 3.

Age verification laws by country: differences that change the decision

The comparison selects jurisdictions with material regulatory differences and institutional activity. It covers two countries in each of North America, Asia, the Middle East, Africa, Latin America and Oceania, plus the EU overlay and three European national systems. EU-wide rules are applied once. The US analysis concentrates on federal law and New York, California, Texas and Georgia; the Canadian analysis includes Quebec. It does not provide every-state, every-province or every-country launch clearance.

United States. The amended COPPA Rule's general compliance deadline was 22 April 2026. Covered operators need the notice, parental-consent, security, vendor-assurance and retention controls. The 25 February 2026 age-verification policy of the Federal Trade Commission (FTC) offers conditional enforcement discretion for general- and mixed-audience services collecting information solely to establish age. Its conditions include purpose limits, vendor diligence, security and prompt deletion. Treat that policy as enforcement policy, with its stated scope and end conditions. COPPA Rule; FTC policy.

For territorial scope, FTC staff explain that covered foreign commercial services directed to US children, or knowingly collecting personal data from US children, fall within COPPA; US-based commercial services can also be covered for children abroad. Apply the rule's operator and commerce conditions to the actual audience and collection. COPPA section 312.2; FTC staff FAQ B7.

State duties change the feature permission:

  • New York. Operative position and trigger: Chapter 368 starts 1 January 2027. Covered unsafe AI features require a permissible determination that the user is 18 or older. Final SAFE for Kids Part 700 rules start 25 January 2027, creating an interim period under the AI law. Product consequence: Classify the enumerated behaviours. Parent consent cannot authorise unsafe features for a minor. The AI cross-reference to permissible methods does not automatically import every exemption applying to addictive-feed platforms. Chapter 368; final rules.
  • California. Operative position and trigger: SB243 companion safeguards, in force as at 19 November 2026, are followed by SB1119's 2027 age route and July 2027 substantive child safeguards. Amended AB1856 device-age provisions become operative 1 January 2027, with a later legacy integration deadline. Product consequence: Treat age signals, safeguards and migration as separate requirements. Applying specified child protections to all users can be an express route under the companion law, subject to its conditions. SB243; SB1119; AB1856.
  • Texas. Operative position and trigger: Chapter 129B covers specified sexual-material websites and a separate AI-tool branch. The app-store law's injunction was stayed by the Fifth Circuit in June 2026; the Supreme Court declined to vacate that stay in July 2026. Product consequence: Check both the content/generation provision and developer/store duties. Facial estimation by itself is not established as a listed Chapter 129B verification route. Chapter 129B; Fifth Circuit; Supreme Court docket.
  • Georgia. Operative position and trigger: Act 518 starts 1 July 2027. Before access to a covered companion feature that may generate sexually explicit conduct, use commercially reasonable assurance proportionate to risk. Product consequence: The act expressly contemplates estimation, account-based assurance and identity verification where necessary. Match the companion definition and feature, and contract for its age-only use and short retention rules. Signed SB540, section 39-5-6(j).

California's September 2026 enactments also matter beyond classic companions. AB1709 is enacted to take effect on 1 January 2027. A covered platform must have addictive features as a significant part of its service, subject to the statutory exclusions. For California residents under 16, the law prohibits those features and permits accounts without them; it contains no parental-permission override. A challenge was filed on 4 November 2026. AB1709; filed complaint.

AB2246 also has a regular 1 January 2027 effective date. It replaces the earlier child-design law for covered California Consumer Privacy Act (CCPA) businesses providing services likely to be accessed by children, with risk-appropriate age estimation or the specified child privacy/data protections for all consumers. Its age-signal cross-reference points to section 1798.501(b); AB1856 places developer signal duties in (d), leaving (b) as an OS sharing restriction. That enacted mismatch requires legal interpretation before relying on the particular signal route. The express all-consumers alternative and other applicable duties remain separately assessable. AB2246; AB1856.

Licentium's existing articles cover New York Chapter 368 and Georgia's July 2027 law in more detail. Their state-specific triggers should feed the common product permission system.

Canada and Quebec. The Personal Information Protection and Electronic Documents Act (PIPEDA) requires appropriate purposes, meaningful consent, safeguards and accountability for outsourced processing. The final May 2026 age-assurance guidance of the Office of the Privacy Commissioner of Canada (OPC) supports a proportionate approach. Under the regulator's child-consent position, under-13s generally require parent/guardian consent; older children still need an understandable consent process suited to the purpose. Quebec separately regulates collection from under-14s and consent by a minor or authorised adult, with a limited clearly-for-the-child's-benefit collection exception. PIPEDA sections 5(3),6.1 and Schedule 1; OPC guidance; Children's Privacy Code; Quebec sections 4.1,14.

Quebec requires an assessment for relevant system development/redesign and before entrusting processing outside Quebec, including another province. Biometric identity or database proposals require the additional statutory and regulator-notification analysis. Canada's C34 and S209 remained parliamentary proposals as at 19 November 2026; neither supplies an operative general AI-account ban. Contract capacity follows the relevant province and transaction. Quebec sections 3.3,17,18.3; CAI biometric guidance; C34 status; S209 status.

Europe: proportionality, effectiveness and national content rules

EU overlay. Article 28 of the Digital Services Act (DSA) requires appropriate and proportionate child privacy, safety and security measures for covered online platforms accessible to minors, subject to the Article 19 enterprise exemption and other conditions. Its final minors guidelines distinguish verification, estimation and risk. They reject self-declaration as an appropriate Article 28 assurance method and preserve platform responsibility when an OS or supplier performs the check. Article 28(3) also guards against treating the provision as a general obligation to collect extra personal data simply to identify minors. DSA Articles 19,28; guidelines paragraphs 24–53.

The EU age-verification blueprint provides an implementation model; it does not certify every vendor offering a similar token. The proposed EU Kids Act remained a proposal as at 19 November 2026. The AI Act separately governs specified age-vulnerability exploitation and AI interaction transparency, with provision-specific commencement dates. An age check does not discharge those behaviour and disclosure duties. Commission blueprint; EU Kids Act proposal; AI Act Articles 5,50,113.

United Kingdom. Highly effective age assurance applies through specific online-safety duties, including covered provider pornography and relevant child-protection obligations. Ofcom assesses technical accuracy, robustness, reliability and fairness across the complete process. As at 19 November 2026, the law requires separate analysis of user-to-user, search and provider-content services. The 2026 legislation creates a power for further AI-service regulation; it did not itself turn every standalone chatbot into the same regulated service. OSA sections 3–4,11–12,79–81; Ofcom guidance; Crime and Policing Act section 248.

France. Covered pornography services must satisfy the framework of the Law for Confidence in the Digital Economy (LCEN) and the regulator Arcom, including privacy separation and at least one double-anonymity option within the technical reference. Other-EU-established services require attention to the individual designation procedure. The temporary credit-card transition has expired. The 2026 constitutional decision struck the blanket under-15 social restriction; the pornography duties remain a separate matter. LCEN Articles 10–10-2; Arcom reference; Decision 2026-911DC.

Ireland. The binding Online Safety Code's detailed requirements applied by 21 July 2025 to the relevant designated video-sharing services, subject to specific extensions. Effective assurance must prevent children from normally seeing adult-only video; self-declaration alone fails. Parental controls for permitted under-16 users and accessible complaints are separate obligations. The September 2026 investigation into X examines those controls; an investigation is not a finding of breach. Code sections 12.10,14,16–17; application decision; investigation notice.

Asia: the app-store actor and the commencement date matter

Singapore. The binding App Distribution Services Code addresses designated stores. Its age/range assurance must achieve reasonable accuracy, with the implementation deadline 31 March 2026. An app developer must account for the resulting distribution controls while separately satisfying its own Personal Data Protection Act (PDPA) duties. The guidance of the Personal Data Protection Commission (PDPC) generally allows understandable consent by 13–17-year-olds and requires parent/guardian involvement below 13; it is a capacity-and-consent interpretation, not a universal statutory admission age. Proposed social-account restrictions announced in September 2026 were planned for legislation in 2027. ADS Code paragraphs 17–22; deadline response; PDPC children's guidance; September 2026 policy response.

India. The substantive child provisions of the Digital Personal Data Protection (DPDP) Act are scheduled for 13 May 2027, as confirmed by the official implementation announcement and Gazette instruments. Existing sensitive personal data or information (SPDI) duties and other applicable duties still need attention before that date. Once in force, section 9 requires verified parental consent for child data, restricts detrimental processing and prohibits specified child tracking, behavioural monitoring and targeted advertising, subject to defined exceptions. A persistent companion-memory system therefore needs a processing-purpose assessment, even if its age gate works. Act section 9; commencement notification; final Rules; November 2026 confirmation.

Once in force, Rule 10 accepts specified reliable existing/provided identity-and-age details or authorised mapped tokens for checking the adult asserting parenthood. The workflow must still address lawful parental authority and consent. A general assistant cannot assume educational or health exceptions merely because users ask school or health questions. Curated audiovisual catalogues have a separate regulatory and litigation branch. Rules 10–12 and Fourth Schedule; IT Rules.

Middle East: broad child rules and distinct data regimes

UAE. The Child Digital Safety Decree-Law covers platforms operating in or directed to the UAE where children use or encounter services. It separates under-13 data controls from risk-based age verification and general under-18 protection. It became effective 1 January 2026 and gives specified existing persons a regularisation period of up to one year, subject to extension. Determine whether the product falls within that existing-person transition. Decree-Law 26/2025 Articles 3,6–8,18,20.

Resolution 106/2026 separately prohibits under-15 social accounts/full features, imposes special restrictions for 15-year-olds, and expressly prevents caregiver consent from overriding the restrictions. It took effect 30 June 2026 with a 12-month transition. Its broad platform definition includes an algorithmic display/recommendation branch, so the actual AI interface needs analysis. The method provisions reject self-declaration and impose accuracy, privacy, inclusion, anti-circumvention and audit conditions. A provider's claimed approval or licence must be checked. Resolution Articles 1–5,8–9.

Apply the relevant onshore, Dubai International Financial Centre (DIFC) or Abu Dhabi Global Market (ADGM) data regime to each entity and processing operation. DIFC Regulation 10 can add AI-processing governance and high-risk conditions to an assurance system. A free-zone data regime does not by itself disapply national child-access rules. Regulated virtual-asset functionality would require a separate assessment. Federal PDPL; DIFC law and Regulation 10; ADGM regulations; Dubai virtual-assets law.

ADGM requires notification to the Commissioner before processing where a data protection impact assessment (DPIA) indicates likely high risk to individuals’ rights. DIFC requires consultation where risks remain particularly high despite mitigation. Its Article 21(3) permits starting before or during consultation only if there is insufficient time to complete consultation in advance, there is a pressing business need, processing is unlikely to override data subjects’ vital interests, and the Law is complied with throughout; applicable directions must be implemented without delay. ADGM sections 34(5),(7); DIFC Article 21.

Saudi Arabia. The Personal Data Protection Law (PDPL) requires appropriate guardian-validity checks where consent is given for a person lacking capacity. Article 28 restricts copying identifiable official documents to the specified legal/public-authority exceptions; consent alone is not listed. A vendor's standard passport-upload process therefore needs a specific legal basis for copying. The capacity analysis also requires care with the Civil Transactions Law's 18-Hijri-year rule. PDPL Articles 2,8,28; Implementing Regulation Articles 11,13,17; Civil Transactions Law Article 12.

For overseas processing, establish destination adequacy or a qualifying transfer exemption, with its applicable safeguards and prior assessment. Article 4(2)(b)'s non-recurring or limited-period route also requires a limited number of data subjects, the specified safeguards and non-sensitive data. The central-operations route in (c) requires necessity and a controller belonging to a multinational group. The direct service/benefit route in (d) requires recipient accreditation and non-sensitive data. Standard contractual clauses work only within a route that permits them and whose other conditions are satisfied. Check the actual destination, purpose, recipient, data classification and onward access before appointing an overseas verifier. Transfer Regulation Articles 3–7.

Africa: child-data evidence and a narrow approved-content route

Kenya. Data Protection Act section 33 already requires appropriate age-verification and consent mechanisms for children's data, considering technology, processing scale, child use and possible harm. Obtain parent/guardian consent and advance the child's rights and best interests, subject to the specific counselling/child-protection exception. The final 2025 guidance of the Office of the Data Protection Commissioner (ODPC) supports proportionate implementation. A general companion offering occasional supportive answers does not establish the exclusive-service exception. DPA section 33; final child-data guidance.

High-risk child/biometric processing can require the statutory impact-assessment and submission process. Vendor contracts need the General Regulations' instructions, confidentiality, audit, deletion and subprocessor controls. The controller's qualifying 72-hour breach deadline and the processor's separate notification duty must be reflected in the operational schedule. The 2026 Online Safety Bill remained a proposal as at 19 November 2026; existing section 33 is the immediate statutory basis. DPA sections 31,42–43; Regulations 24–25,49–52; Parliament bill.

South Africa. POPIA sections 34–35 restrict children's information and provide defined exceptions, including prior consent by a competent person. Establish who has the relevant legal authority. The platform remains accountable for its operator and needs a written security arrangement and compliant transfer route. POPIA sections 8,19–21,34–35,57,72; regulator guidance.

The online X18 film/game route is much narrower: registered distribution and Film and Publication Board (FPB) exemption/approval, child exclusion and additional controls are required. It includes a one-year register of user name, address and verifiable age. That rule defeats a universal promise to retain only an anonymous token, but does not require storing every raw ID image. Its card-related provision concerns promotions within that regulated route. Amendment section 23 / Act section 24(3); Regulation 28.

Latin America: platform duties reach different AI features

Brazil. The Digital Statute of Children and Adolescents (ECA Digital) has applied since 17 March 2026 to covered child-directed or likely-access products/services. Article 9 requires effective, reliable age verification at each access to content, products or services whose offer or access is improper, inappropriate or prohibited for under-18s, and rejects self-declaration. Decree 12.880 expressly addresses natural-language AI and treats specified sexual/erotic automated dialogue as pornography. The legally defined protected-version routes require actual content and feature separation. ECA Articles 1,9,41-A; Decree Articles 11,16,19,22.

OS/store signals support a wider chain of duties. The product supplier must still adapt and protect its own service, including the relevant browser-access route. A conflict between its assessment and the OS/store signal calls for the more protective result. Documentary inputs must be eliminated irreversibly after the required age information is captured under Decree Article 24(3). Technical implementation work by the National Data Protection Authority (ANPD) remained in progress as at 19 November 2026; an unnamed vendor should not be described as certified. ECA Articles 12–15; Decree Articles 24–30; ANPD agenda.

Article 24's parent-account language uses “up to 16” and a broad child-directed/likely-access supplier formulation, despite its social-network chapter heading. Assess that wording and scope question rather than reducing it to under-16 social accounts. Privacy consent and best interests must separately satisfy Article 14 of the General Data Protection Law (LGPD) and the applicable legal basis. ECA Article 24; LGPD Article 14; ANPD interpretation.

Mexico. The 14 October 2026 child digital-safety decree is operative. For platforms within its intermediary/inter-user definition, Article 4(I) requires reliable, secure age-identification mechanisms beyond self-declaration, alongside parental tools, child-safe defaults and periodic risk assessment. A strictly private standalone bot requires a reasoned classification against the between-users wording. Article 3's harmful-content restrictions contain no parent-override route. Decree Articles 1–4 and transitory provisions.

Apply the 2025 private-sector data law and the relevant representation/capacity rules. Its purpose, proportionality, security, rights and transfer requirements constrain any identity check. The decree's minimum 90-business-day preservation concerns removed account/content evidence; it should not be copied into a universal raw-ID retention term. LFPDPPP; decree Article 5; Federal Civil Code.

Oceania: social accounts, restricted generation and facial categorisation

Australia. The under-16 social-account regime and the restricted-content codes require separate classifications. Apply the social definition together with the March 2026 rules' additional feature test. A private assistant is not covered solely because its conversation feels social. AI companions and high-risk generation can nevertheless engage the separate Designated Internet Services code. Its Tier 1 pre-access requirements and Tier 2 generation-prevention/testing route differ. Online Safety Act sections 63C–63F; social rules; DIS Code sections 3–4 and Table 10.

The restricted-content app-distribution measure 1(b) reached its separate 9 September 2026 deadline, subject to documented exceptions and alternative action. Government-ID use also has specific constraints within the social age regime. Document the actual service, code tier, content risk and permitted method. Privacy consent/capacity and parent tools remain separate; parental agreement does not waive the covered social-account minimum. App Distribution Code; Online Safety Act; OAIC age-assurance guidance.

New Zealand. The Biometric Processing Privacy Code can apply to facial age categorisation even when the system is not naming the user. The existing-processing transition ended 3 August 2026. Necessity, alternatives, effectiveness, proportionality and rights must be assessed; an age-categorisation exception does not remove the whole code. Biometric code.

Classification law restricts supplying R18 material to minors. The Classification Office distinguishes that prohibition from a mandated online age-verification system. Bill 339–1's proposed social and AI-companion scheme remained a proposal as at 19 November 2026, with commencement dependent on Royal Assent. Apply the privacy law and contract-capacity rules to the actual user and transaction while monitoring that separate bill. Classification Act; Classification Office; bill status; Privacy Act; Contract and Commercial Law Act.

Which age-assurance method is legally sufficient?

Assess the complete process: evidence source, capture, binding to the user, decision threshold, uncertainty, circumvention resistance, privacy, integration and review. A vendor's single accuracy percentage cannot answer those questions. Ofcom's guidance and report, the EU guidelines and the UAE/Brazil frameworks require attention to different aspects of effective operation. Ofcom guidance and 2026 report; EU guidelines; UAE Article 4; Brazil Article 24.

  • Self-declaration. Evidence to obtain from the supplier: Exact question, routing, treatment of conflicting information and why the applicable rule permits it. Limits that affect legal sufficiency: Insufficient alone for UK highly effective assurance, French pornography, Irish adult-video controls, Brazilian restricted content, Mexican covered-platform duty and UAE social method rule.
  • Document check. Evidence to obtain from the supplier: Issuer/document validity, authenticity, holder binding, lawful copying, returned fields and actual deletion. Limits that affect legal sufficiency: The document must relate to the user. Copying and retention have local restrictions, including Saudi Article 28 and Brazilian immediate deletion after extraction.
  • Facial age estimation. Evidence to obtain from the supplier: Results near each relevant boundary, false adult/minor classifications, inconclusives, subgroup performance, spoof resistance and model version. Limits that affect legal sufficiency: A probabilistic result needs an uncertainty policy and alternative. Biometric definitions differ. A method allowed in one country may not fit another country's statutory list.
  • Identity credential or age token. Evidence to obtain from the supplier: Issuer authority, underlying method, authentic assertion, account/user binding, expiry, correction, revocation and tracking resistance. Limits that affect legal sufficiency: A token derived from self-declaration has different evidence from a verified credential. An account-linked token can remain personal data.
  • Financial or mobile-network signal. Evidence to obtain from the supplier: Whose age was checked, legal minimum at issuance, source freshness, current user binding and limited lawful access. Limits that affect legal sufficiency: An adult-owned card or device can be shared. COPPA's payment-consent route has specific transaction and notice conditions. Debit cards do not establish adulthood.
  • Email, account or behavioural inference. Evidence to obtain from the supplier: Lawful source data, model validation, timing and evidence that the resulting certainty meets the actual duty. Limits that affect legal sufficiency: Inference after use cannot protect a pre-access adult gate. Profiling or child tracking may itself be restricted.

The method-specific limits above follow the Ofcom guidance, COPPA consent rule, French reference, Irish Code, Brazil decree, Mexico decree, UAE Resolution and Saudi PDPL. They establish conditional tests rather than a global approved-method list.

Age verification and GDPR: collect the attribute the decision needs

For an 18+ restriction, first ask whether a reliable over-18 assertion meets the legal standard. Exact birth date, name, passport number and a reusable facial template each need a separate justification. Document the lawful basis, any special-category condition, necessity, retention and transfer route. The age-assurance statement of the European Data Protection Board (EDPB) supports proportionate design that limits identification and cross-service tracking. GDPR Articles 5–6,9,25,35,44; EDPB Statement 1/2025.

Under EU GDPR, the special biometric category turns on processing for unique identification. New Zealand's code can reach age categorisation, while Brazil's sensitive-data wording requires its own assessment. Determine the technical operation in each regime. A supplier's “non-biometric” marketing label cannot substitute for that analysis. GDPR Articles 4,9; New Zealand code; LGPD Articles 5,11.

Three workable designs to compare

  • 1. Protected general assistance with restricted features locked. When it can work: The product enforces child-protective behaviour and data defaults; restricted generation, unsafe relationship features and prohibited tracking remain unavailable. The relevant child-data route is lawful. Evidence and residual issues: Demonstrate prevention and leakage testing, protected defaults, the consent/authority route and separate permissions. A mandatory age-identification or OS-signal duty can still apply. California's express all-user routes, Brazil's protected-version conditions and UK child privacy standards each have defined scope.
  • 2. Minimal credential or age-attribute result. When it can work: A reliable/authorised issuer performs the required check and returns the minimum necessary threshold or range through a secure account-bound process. Evidence and residual issues: Verify underlying evidence, issuer status, freshness, replay protection, revocation, privacy separation and alternatives. France's double-anonymity requirement and India's 2027 authorised-token route require their particular conditions. Store-age signals have jurisdiction-specific legal treatment.
  • 3. Calibrated estimation with a lawful escalation route. When it can work: Estimation is permitted and empirical evidence supports the actual age boundary and user population; uncertain cases receive another usable method. Evidence and residual issues: Validate false results and inconclusives near the boundary, accessibility, spoof resistance, privacy basis, deletion and appeals. Check restrictive method lists and document-copy rules before offering escalation. A person should not be forced into an unlawful ID-copying process to challenge an estimate.

These alternatives draw on California's companion provisions, child-design provisions, Brazil's decree, the child standards of the UK Information Commissioner's Office (ICO), French reference, India Rule 10, Ofcom method criteria and Saudi Article 28. Select a design per feature and territory; shared infrastructure should preserve those legal branches.

Put an enforceable age-assurance schedule in the supplier contract

The schedule should connect the supplier's promised performance to the platform's actual duty. A general security annex or a statement that the service is compliant leaves essential questions unanswered.

  • Scope and roles. Required commercial and technical evidence: Features, countries, age boundaries, purposes, inputs/outputs, controller/processor allocation and any supplier-owned purpose.
  • Method and performance. Required commercial and technical evidence: Exact deployed model/settings, evidence source, threshold-specific error and inconclusive results, representative populations, anti-circumvention work, independent test scope and required certification.
  • Permitted use. Required commercial and technical evidence: Express limits on training, advertising, identity enrichment, resale, cross-customer matching and telemetry. Test actual logs and model pipelines against the promise.
  • Retention and deletion. Required commercial and technical evidence: Separate schedules for raw captures, extracted attributes, decision records, consent/authority evidence, security logs and incident preservation; deletion across backups and subprocessors.
  • Security and incidents. Required commercial and technical evidence: Access, segregation, encryption and response evidence; supplier notice early enough for the platform's applicable deadline; investigation and regulator assistance.
  • Subprocessors and transfers. Required commercial and technical evidence: Named locations/recipients, authorisation and change rights, flow-down, lawful transfer route, remote access and onward-transfer controls.
  • Rights and errors. Required commercial and technical evidence: Alternative methods, correction propagation, parent withdrawal, significant automated-decision safeguards and usable escalation.
  • Audit and changes. Required commercial and technical evidence: Evidence access, regulator cooperation, notice before material method or purpose changes, retesting and ability to suspend an unreliable route.
  • Liability and exit. Required commercial and technical evidence: Lawful loss allocation, appropriate remedies and verified insurance if relied on; portable permitted results, continued correction/revocation and proven return/deletion.

Some terms are expressly required in processor law; others are proposed mechanisms for demonstrating the required outcome. GDPR Article 28, Kenya Regulations 24–25, DIFC Article 24, ADGM section 26, Saudi implementing Article 17 and Quebec section 18.3 provide concrete contractual anchors. Ofcom's vendor guidance explains the performance evidence. An indemnity allocates private losses within its lawful scope; it does not relieve the platform of an access or data duty imposed on it.

Retention needs a country-and-purpose schedule

The correct retention answer can be prompt deletion, a short outside limit or a specified minimum record. Apply each rule to the data it actually covers.

  • Texas Chapter 129B. Consequence: The regulated entity and verifier may not retain identifying information under the applicable provision. Section 129B.002(b).
  • Brazil documentary age checks. Consequence: Eliminate the document image/copy/information irreversibly once the necessary age information is captured. Decree 24(3).
  • Georgia, from 1 July 2027. Consequence: Use age-assurance data only for age verification and retain it no longer than 24 hours, or another specified period permitted by law if longer. Identity documents may be retained only as reasonably necessary to complete assurance unless another law requires retention. Minimise collection and retention. Section 39-5-6(j).
  • South African approved online X18 route. Consequence: Keep the specified name/address/verifiable-age access register for one year. Raw image retention needs its own justification. Act 24(3); Regulation 28.
  • India substantive Rules from 13 May 2027. Consequence: Rule 8(3) covers personal data, associated traffic data and processing logs, including processing by a processor, for Seventh Schedule purposes: retain them for at least one year from the processing date, then erase unless further retention is required by another law or notified by government. Map this duty and Rule 6 security retention before promising a purge. Rules 6,8.
  • Mexico child-platform decree. Consequence: Minimum 90-business-day preservation relates to removed account/content evidence. Article 5.
  • Quebec decision information. Consequence: Keep personal information used to make a decision about a person for at least one year after that decision. Identify the data actually used and reconcile applicable biometric-specific deletion rules. Section 11.
  • New York final Part 700, where directly applicable. Consequence: Input-deletion duties coexist with separate five-year compliance and ten-year testing/certification record requirements. The AI law's method cross-reference needs its own scope analysis. Sections 700.5,700.7.

Legal holds should identify the rule, fields, purpose, access controls and end condition. They should not become a general exception permitting the vendor to keep all inputs indefinitely.

Licentium's Texas biometric analysis for a Brazilian KYC supplier addresses that particular biometric and transfer corridor. For a consumer age-check project, first establish whether the supplier performs identification, age estimation or both.

Parents, mistakes and app stores need their own operating procedures

A parent workflow should establish the consenting person's legal authority, identify the particular processing or feature permission, record informed consent and support withdrawal. Privacy consent cannot open a feature that the applicable law reserves for adults or otherwise prohibits for children. Kenya, Saudi Arabia and EU GDPR expressly illustrate the importance of authority; New York, UAE social rules and restricted-content laws show the limits of parental permission. Kenya section 33; Saudi Article 13; GDPR Article 8; NY section 1801; UAE Article 3. Apply any separate age or capacity condition. New York's final feed rules expressly recognise a parent who is themselves a minor; India's Rule 10, once in force, requires the person identifying as parent to be an identifiable adult. NY section 700.1(cc); India Rule 10.

For mistakes, provide a clear route to correct age/account data, submit proportionate alternative evidence and obtain meaningful review where required. Apply the relevant automated-decision test. UK Articles 22A–22C differ from EU Article 22, so a copied global rights notice can be inaccurate. Keep an adult-only feature protected during uncertainty, and propagate a justified correction across the account, supplier and permitted store signals. EU GDPR Articles 16,22; UK GDPR Articles 22A–22C; Brazil decree Article 27.

As at 19 November 2026, Apple's “confirmed” age-range result can combine underlying confirmation routes; inspect the documented semantics before claiming a particular method was used. Google's Age Signals API has regional/state-dependent availability and purpose restrictions, including prohibitions on advertising, marketing, profiling and analytics uses. An unavailable or unknown response supplies no basis to assume adulthood. California's amended 2027 statute also restricts requesting signals without the relevant legal requirement. Apple support; Google API documentation; AB1856.

Store review policy can independently prevent a proposed distribution. Apple's rules exclude specified pornographic content and impose safeguards for generated/user content. Its parental gate is distinct from privacy consent. Google has its own generated-content and Families requirements. Assess the actual app and current agreement before promising availability in a store. Apple Guidelines 1.1.4,1.2,5.1.4; Google policies.

Launch checklist and documented rechecks

Before release, assign an owner and evidence for each gate:

  1. Classify the feature and nexus. Record the countries/states, user groups, public-sharing, companion behaviours, content modalities and applicable effective dates.
  2. Define the permission. Keep access, data consent, parent authority, contractual capacity and distribution approval separate.
  3. Complete privacy prerequisites. Approve the data flow, roles, bases, biometric assessment, required DPIA/filing, retention and transfer route.
  4. Validate the method. Obtain deployed-system evidence at the relevant boundaries, including errors, uncertainty, subgroup performance, spoofing and user binding.
  5. Sign the supplier schedule. Resolve reuse, subprocessors, incidents, audit, deletion, changes, liability and exit with the actual vendor.
  6. Test every entry route. Cover direct web, mobile, unsupported/older OS versions, signed-out access, shared devices, new and existing accounts, fallback and appeals.
  7. Verify distribution. Check ratings, policy content restrictions, API permissions, parent download/update requirements and the actual developer contract.
  8. Prepare operations. Make corrections, withdrawals, reviews, incidents and deletion work in practice; assign a legal-change owner.

These gates implement the duties identified above, including GDPR accountability and processor controls, Kenya's assessment/contract rules, Ofcom's effectiveness evidence, Brazil's supplier-chain rules and the distribution documentation. They are a proposed management checklist, not a claim that every jurisdiction prescribes one identical form.

Recheck when credible evidence conflicts with age status, the user changes device/account control, parental authority or consent changes, a threshold is reached, a risky feature is added, the supplier changes a model or purpose, or a law or court order changes. Respect express repeat-check requirements, including Brazil's restricted-access rule. For other cases, document why the trigger or interval is proportionate. Retain the applicable legal version, result, reason, method/version, uncertainty, permitted authority record, review outcome and deletion state, with only necessary fields.

The 2027 calendar includes New York's 1 January 2027 AI rule and 25 January 2027 final method rules, California's staged 2027 changes, India's 13 May 2027 substantive DPDP phase, Georgia's 1 July 2027 law and the UAE social transition anniversary on 30 June 2027. EU Kids Act, Canadian C34/S209, New Zealand Bill 339–1 and announced Singapore/UK social measures remained proposal or future-implementation items as at 19 November 2026.

Review the age-assurance design

For a scoped review, assemble the feature map, launch-country list, age and parent flows, data diagram, supplier testing and proposed contract, retention schedule and store implementation. Licentium's Data Protection & GDPR for AI page describes related work on lawful bases, data flows, impact assessments, processor terms, transfers and retention.

Planning a launch somewhere new?

Prefer email? Write to hello@licentium.io. By sending, you accept our privacy policy.