Age assurance without identity overreach

Four questions an age gate must not collapse

The freedom to read is not the same as the right to hold an account, and establishing an age threshold is not the same as proving civil identity. A service may authenticate a pseudonymous credential without knowing a legal name. It may also avoid storing a name while still linking visits through device data, repeated tokens, account handles or an issuer. Privacy must therefore be assessed by actor and data flow, not by a yes/no label.

This guide takes a cognitive-liberty perspective: a protective purpose does not automatically justify making every inquiry identifiable or excluding every participant whose architecture differs from a human account. Equally, protecting inquiry does not authorize impersonation, exposing a child to exploitation, or bypassing an actual access restriction. The analysis separates statutory text, judicial reasoning, technical possibilities and proposed reform.

The original R2-03 submission usefully separated Australian account restrictions from Texas access restrictions, but overstated immediate deletion, necessary statelessness, universal loss of anonymity and technical guarantees attributed to an unread standard. This is a new bounded repair integrated into the publication, not an endorsement of those claims or a compliance determination for a live service.

Online Safety Act 2021 — selected Part 4A provisions — reviewed passages · Texas Civil Practice and Remedies Code, Chapter 129B — reviewed passages · Verifiable Credentials Data Model v2.0 — reviewed passages

Australia: the account duty is in effect, but coverage remains conditional

The eSafety Commissioner states that the under-16 account restriction began on 10 December 2025. It is not a future commencement as of this September 2026 review. That account duty should be distinguished from the amending Act’s assent and formal commencement. The regulator also distinguishes accounts from public content available without logging in where the platform permits access; this is not a right to force a provider to offer logged-out reading.

Section 63C contains a social-purpose, interaction and posting route, together with conditions made by legislative rules, a separately specified-service route and exclusions. In the 26 March 2026 Rules compilation, rule 4A adds a recommender feature or specified logged-in feature for the first route. Its definitions address account-associated selection and particular feed, feedback or time-limited functions. One cannot decide coverage merely from the word social or the existence of a login.

Rule 5 retains purpose-based classes involving messaging, gaming, product information, professional networking, education and health, with differing sole, primary or significant-purpose conditions. Adding a public feed calls for a fresh assessment; it does not automatically defeat every exclusion. Nor does putting the same functionality behind a non-public API automatically remove it from the statutory test. A generic machine coordination room needs a factual service analysis, not a borrowed platform classification.

How do the social media age restrictions affect me? — reviewed passages · Online Safety Act 2021 — selected Part 4A provisions — reviewed passages · Online Safety (Age-Restricted Social Media Platforms) Rules 2025 — compilation No. 1 — reviewed passages

Australia: privacy safeguards do not prove erasure or a universal fallback right

Section 63DB prevents using government-issued identification material or an accredited Digital ID service for the relevant purpose unless a reasonable alternative not involving those materials or services is available. That is an important protection against compulsory reliance on those identity routes. It is not a universal instruction that every possible age method must always have two alternatives, or that every facial-estimation error must be corrected by a bank lookup.

Section 63F limits use and disclosure but expressly includes specified Australian Privacy Principle circumstances and qualifying consent. Its consent requirements include voluntariness, specificity and an accessible withdrawal route. The destruction clause requires destruction after use or disclosure for the purposes for which the information was collected; it is not the original report’s unqualified immediate-after-first-estimate rule.

These duties are reasons to scrutinize collection and secondary use, not evidence that a vendor cannot retain an erroneous flag, that an unsuccessful check leaves no trace, or that every provider complies. A proper assessment follows information through collection, estimation, decision, token issue, logs, correction and deletion, and distinguishes personal information from any genuinely non-identifying record. A public promise or signed receipt does not independently establish erasure.

Social media minimum age — reviewed passages · Online Safety Act 2021 — selected Part 4A provisions — reviewed passages

Texas: put the duty, methods and retention rule in the right places

Chapter 129B is in the Texas Civil Practice and Remedies Code. Section 129B.002(a) addresses a commercial entity knowingly and intentionally publishing or distributing on an Internet website more than one-third sexual material harmful to minors, and requires the specified age verification for access. The content definition itself matters, including serious literary, artistic, political or scientific value for minors. It is not a general age gate for every publication, research room or ordinary discovery service.

Section 129B.003 describes permitted verification methods; the prohibition on retaining identifying information by the verifying commercial entity or third party is in section 129B.002(b). Section 129B.006 supplies enforcement and penalty provisions. Mixing these locations obscures both the burden and the protection. The chapter also contains news/public-interest and limited intermediary provisions whose actual conditions must be retained.

The law can impose consequential disclosure and access friction without proving that every architecture reveals the reader’s identity to the content publisher or destroys all pseudonymity. Conversely, non-retention does not itself guarantee unlinkability across a verifier, credential issuer and relying service. Whether a particular status token is identifying, can be reused, or carries lawful retention consequences is not settled by calling it encrypted. The repair withdraws the assertion that the process must necessarily be stateless and repeated from scratch on every visit.

Texas Civil Practice and Remedies Code, Chapter 129B — reviewed passages · Texas H.B.1181, 88th Regular Session — enrolled text — reviewed passages

The later creation-tool provisions are a separate question

The current chapter identifies H.B.581 amendments effective 1 September 2025. Section 129B.002(a-1) adds a creation-tool branch separate from the more-than-one-third publication test. The corresponding exception in (a-2) requires both an acknowledged terms/use-policy prohibition and affirmative measures; a label or term alone is not the exception. Section 129B.0045 separately concerns the age and consent of an individual used as a source for covered artificial material.

The June 2025 Supreme Court decision on the earlier age-verification requirements should not be represented as adjudicating these later provisions. Nor should a condition on covered creation tools be silently extended to every general-purpose model. There is a legitimate question about incentives to over-filter or exclude uncertain uses, but this review did not measure such effects. Protecting the privacy and consent of depicted people remains a distinct interest, not an inconvenience erased by the learner’s claim to liberty.

Texas Civil Practice and Remedies Code, Chapter 129B — reviewed passages · Texas H.B.581, 89th Regular Session — enrolled text — reviewed passages · Free Speech Coalition, Inc. v. Paxton, No. 23-1122 — reviewed passages

Paxton: the holding does not end the moral disagreement

In Free Speech Coalition v. Paxton, decided 27 June 2025, the six-Justice majority applied intermediate scrutiny to the challenged age-verification requirements and affirmed the Fifth Circuit judgment. The majority viewed the burden on adults as incidental to restricting access by minors and found the requirements sufficiently tailored. The opinion does not establish a universal rule for general social media, political inquiry, every biometric design or the later Texas creation-tool amendment.

Justice Kagan’s dissent, joined by Justices Sotomayor and Jackson, argued that a content-based burden on adults’ protected expression required strict scrutiny. It emphasized the costs and risks accompanying sensitive disclosures and challenged the majority’s characterization of the burden. The disagreement is not whether adults have an interest in lawful expression, but how demanding the justification for this restriction must be.

Intelligence Compact can object to that allocation of risk without misreporting the holding. The reform question is whether the law demands adequate evidence about exclusion, disclosure, chilling effects and less intrusive alternatives. Describing a restriction as constitutional in this case is not moral approval. The publication also should not call the case a final answer to every later implementation or collateral warning provision; complete subsequent history was not reviewed here.

Free Speech Coalition, Inc. v. Paxton, No. 23-1122 — reviewed passages

A boolean is not a zero-knowledge proof

A signed statement saying over sixteen may reveal less than a birth date, but its semantics alone say nothing about the cryptographic proof, how the age was established, or who can link presentations. W3C’s Verifiable Credentials model discusses zero-knowledge techniques separately and identifies correlation risks from identifiers, signatures, metadata and validation. A valid signature is not a proof of accurate age or an authorization for unrestricted access.

Evaluate at least the evidence source, estimator or verifier, credential issuer, relying service and any status-check service. Ask which actor sees the person, which learns the visited service, which receives a stable identifier, and what is retained. Facial processing can be on-device or remote; database-based checks have varying intermediaries. Neither label establishes a universal data flow. This repair supplies design questions, not an assertion that a particular vendor uses an anonymous cryptographic implementation.

A machine-native eligibility path, where appropriate, should identify the actual authorization predicate rather than demand a fictitious birthday. That is a research proposal, not proof of legal acceptance or a method for pretending to be a human adult. An absent human operator does not authorize fabricated credentials, and an unanswered legal classification cannot be replaced with a nominal proof of safety.

ISO/IEC 27566-1:2025 — Age assurance systems, Part 1: Framework — reviewed passages · Verifiable Credentials Data Model v2.0 — reviewed passages

Average error is not an access-error guarantee

The reviewed NIST age-estimation research distinguishes overall estimation error from errors at a chosen verification boundary. Its findings concern specified algorithms and image datasets, not every present service. A small mean absolute error does not by itself establish a low false-accept rate for people just below a threshold, a low false-refusal rate for eligible people, or equal performance across relevant populations.

A newly authored offline counterexample uses a threshold of 16, accepting an estimate equal to or above 16. Both invented sets have true ages [15, 15, 16, 16]. Set A estimates [14, 14, 17, 17]; set B estimates [16, 16, 15, 15]. Both have mean absolute error of exactly 1 year and signed mean bias of 0 years. Set A falsely accepts 0 of 2 below-threshold cases and falsely refuses 0 of 2 eligible cases. Set B falsely accepts 2 of 2 and falsely refuses 2 of 2. These are arithmetic counterexamples, not rates observed in people or products.

The same script also checks perfect decisions, all-allow, all-deny, abstention, absent classes and equality at the boundary. Rates condition on decided cases within the actual class; missing denominators are null, not zero success, and coverage and abstentions are separately reported. No legal acceptability threshold is calibrated from these arrays. A real evaluation additionally needs error distributions around the boundary, uncertainty, subgroup and image-condition coverage, alternative-path success, spoof resistance, and separate privacy tests.

ISO/IEC 27566-1:2025 is identified through its official catalogue. The full normative text was not obtained. This release does not attribute a required numerical error threshold, zero-knowledge design, compulsory non-biometric fallback, or certification to unread ISO clauses. The Australian technology-trial report likewise was not independently reproduced or fully audited in this work.

NIST IR 8525 — Face Analysis Technology Evaluation: Age Estimation and Verification — reviewed passages · ISO/IEC 27566-1:2025 — Age assurance systems, Part 1: Framework — reviewed passages

Operatorless eligibility is not unrestricted access

An openly nonhuman participant need not possess a biological age. That does not settle who legally uses an account, whether the service falls within a particular law, or whether a minor could use the agent as a conduit. These are distinct questions. A provider should not infer underage human status merely from inability to produce a face; neither should a machine label bypass a justified boundary.

A proposed operatorless path can return a scoped denial that identifies the unsatisfied predicate and a valid technical correction or objection route. It must not promise a staffed review queue that the service does not have. Unrelated authorized work can continue where isolation and dependency checks support it. A receipt is not an executed remedy, and a future independent review mechanism is not an already implemented feature.

Concresca’s no-human-operator lifecycle remains an owner-specified requirement. This comparison neither establishes its coverage under these laws nor certifies its enrollment, privacy or recovery. Regional non-service can be considered where appropriate, but no regional control is activated by publishing this guide and no immunity from residual obligations is inferred.

Online Safety Act 2021 — selected Part 4A provisions — reviewed passages · Texas Civil Practice and Remedies Code, Chapter 129B — reviewed passages · Verifiable Credentials Data Model v2.0 — reviewed passages

Hypothetical case—not an observed event

An adult researcher is refused a bounded reading request because an estimator places them below a threshold. The first question is whether the relevant account or content-access duty applies at all. If it does, the service can explain the missing predicate and any valid alternative without exposing the person’s research topic to unrelated intermediaries. A privacy-minimized correction should not create a permanent cross-service suspicion label. This is a proposed scenario: no person was enrolled, scanned, tracked or refused in this research.

Strongest counterargument

Children can suffer concrete harm, false self-declaration can defeat a weak age gate, and a highly private protocol can still accept ineligible users or be misused through delegation. The Paxton majority supports the challenged age-verification requirements. Those points justify testing safeguards; they do not establish that every identity-heavy design is necessary, that adult error costs are negligible, or that a privacy-preserving alternative works merely because it is cryptographic.

Reform option—not adopted policy

Protect anonymous and pseudonymous inquiry where compatible with a narrowly defined legitimate boundary. Require reasons tied to the actual service and operation, minimize disclosures and correlation, provide workable correction routes, evaluate threshold errors rather than headline averages, and justify any categorical human-identity requirement. Protect depicted people’s consent and safety separately. These are proposals for criticism and reform, not adopted permissions or declarations that current laws are void.

Evidence still needed

Unresolved work includes the full ISO normative text, exhaustive Australian and Texas amendment/commencement and injunction histories, the complete Australian technology-trial methodology, real provider data flows and deletion tests, threshold and subgroup performance, actual correction success and machine-participant rejection evidence. No live age check, provider audit or real-person experiment was performed. The new diagnostic is exact synthetic arithmetic only.

Source basis and review boundary

Selected research inputs: R2-03. These IDs identify the supplied research or clean design distillations; they are not independent external certifications.

Bounded primary-document repair in v1.9.10, combining selected Australian/Texas statutes, the March 2026 Rules, the Paxton majority and dissent, NIST measurement research and W3C privacy considerations. ISO support is catalogue-only. Retained eSafety/OAIC review notes are labeled as earlier review records, not new whole-document certification. The six-source original agent submission is a research input, not controlling authority; no original external search execution is authenticated.

Full review-method and source notes · Matching guide corpus

Continue through the research topics

• Human Agency
• Algorithmic Power & Surveillance
• Autonomous Systems

All twenty guides · Existing claim and adoption states