Two things are happening on the same timeline and almost nobody is discussing them together.

One: governments across the US, UK, Australia, and EU are mandating age verification for pornography, social media, app stores, and increasingly general-purpose websites. Texas’s App Store Accountability Act is in force after the Supreme Court declined to block it. The UK’s Online Safety Act duties are live. Australia’s under-16 ban is a year old. The Senate has advanced four bills that require identifying internet users.

Two: the technology those mandates depend on is being destroyed by generative AI, in public, right now.

The economics collapsed

A convincing forged identity document can now be produced for roughly $15 in about half an hour. What previously required specialist equipment, template stock, printing knowledge, and physical craft is now a prompt and an image model.

The Resistant AI Global Document Fraud Report 2026 puts GenAI-driven document fraud detections up 90% year over year — the fastest-growing fraud vector in identity verification.

The DOJ’s OnlyFake case shows the industrial version: a service that generated over 10,000 fake digital IDs spanning 50 US states and 56 other countries, sold for cryptocurrency in bulk packages of up to 1,000 at a time, explicitly marketed for bypassing KYC. Its operator pleaded guilty this year. The service is gone; the capability it demonstrated is not.

Why the checks fail

The failure is architectural, not a matter of better models.

Upload-only flows accept static files. You submit an image. An attacker can iterate infinitely at near-zero cost, resubmitting until one passes. Any check that scores a static artefact loses to an adversary who can generate unlimited artefacts and observe the pass/fail signal.

Template, OCR, and database checks confirm format, not authenticity. They verify that a document looks like a Michigan driver’s licence and that the fields parse. A generative model that has seen many Michigan driver’s licences produces something that looks like one.

Liveness checks are being defeated in composite. The sophisticated attack chains a synthetic document, a deepfake portrait matching that document, and a synthetic identity built from a mix of real and fabricated data with a plausible history. Each layer is credible in isolation. Together they defeat systems designed to verify one element at a time.

Which is the essential point: identity verification was built to check individual artefacts against individual sources. Generative AI produces internally consistent sets of artefacts. The architecture assumed forging one thing was hard and forging everything consistently was harder still. That assumption is now false.

The asymmetry that makes it worse

Here is the part that should determine policy and does not.

A motivated fifteen-year-old will defeat an age check. Teenagers are the most motivated, most technically fluent, most peer-networked population imaginable for this problem. The workaround — a VPN, a borrowed credential, a generated document, or one of the several free services that will do it for them — propagates through a school in days. Australia’s regulator found more than eight in ten under-16s still using social media despite the ban, and platforms allowing children to retake age assurance tests until they got an over-16 result.

Meanwhile, every ordinary adult complies. They upload their driver’s licence, or submit to a face scan, or hand over the last four of their SSN, because they are not motivated to route around it and they want to read the article.

So the measured outcome is: the people the law targets evade it, and the people it does not target build a permanent, breach-exposed record of their identity attached to their browsing.

That is not an implementation problem to be fixed with a better vendor. It is the equilibrium the design produces.

The database nobody wants to be responsible for

Age verification mandates create honeypots by construction.

Every site subject to a mandate must either collect government ID directly or route users through a third-party verifier. Direct collection means thousands of websites of wildly varying competence holding scans of driver’s licences and passports. Third-party routing means a small number of verification providers accumulating the highest-value identity dataset ever assembled, linked to what each person was trying to access.

We already know how this ends. Research presented at Black Hat and DEF CON mapped the attack surface. The AU10TIX exposure, the Tea app breach, and a steady procession of verification-vendor incidents have already leaked exactly this category of data.

And the ACRO reprimand this week is the reminder of the base rate: a UK government body left a public-facing portal holding passport scans and biometrics unpatched for four years. That is the average standard of care you should assume across thousands of newly-mandated collectors.

The honest alternatives

There are better designs, and they are not being mandated.

Device-level attestation. The operating system knows, from account setup and family controls, whether the user is a minor, and can assert a signed “over 18” or “under 18” claim to a site without transmitting identity. Apple and Google have shipped versions of this. It leaks a binary rather than a document.

Zero-knowledge age proofs. Cryptographic attestations, issued once by a trusted party, that prove a predicate — this person is over 18 — without revealing who they are or which sites they use. The EU’s digital identity work includes this capability. It is real, deployed, and mature enough to legislate toward.

Neither is a silver bullet: a determined teenager can borrow a parent’s device or credential, and no scheme survives someone handing over an adult’s phone. But that is also true of every ID-based check, at a small fraction of the privacy cost.

The reason legislatures keep mandating document upload instead is not technical. It is that document upload is legible to legislators and produces an auditable record of compliance, whereas cryptographic attestation is not and does not.

What it means in practice

Mandates are arriving as the mechanism collapses. Compliance will be measured by whether a check was performed, not whether it worked.

The cost is borne entirely by compliant adults. Evaders pay $15 and thirty minutes. Everyone else pays with a permanent identity record.

Verification vendors are becoming systemically important, unregulated infrastructure. A handful of companies are being handed the identity documents of entire national populations by statute, with no equivalent statutory security standard.

Expect the ratchet. When the first round of checks demonstrably fails, the response will not be to abandon the approach. It will be to demand stronger checks — biometrics, liveness, continuous re-verification. The failure of a privacy-invasive measure is normally used to justify a more invasive one.

What you can do

  1. Prefer device-level or OS-level age signals where a site offers a choice. Some now offer “verify with your device” alongside document upload. Take it every time.

  2. Never upload a document to a site that is not the one you are trying to access. If a redirect to a verifier happens, check the domain. Fake verification flows are already an established phishing pattern, and they are extremely effective because the real flow looks identical.

  3. If you must upload, use a passport rather than a driving licence where possible. A passport does not carry your home address.

  4. Watermark what you upload. Overlay text stating the date and the specific purpose across the image. It does not stop a breach; it makes reuse of the scan for account opening substantially harder.

  5. Support age-assurance approaches that transmit a yes/no rather than a document. The technology exists. The reason it is not mandated is that nobody is asking for it — see the wider age verification fight.