Product1 distinct publisher3 min readPublished
The longest section of the deal is about telling adults from children, and both available techniques collect more from everyone: behavioural inference on accounts already being logged, or a face scan before you can send a message.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
A Roblox player who wants to send a message to another player points a phone at their own face, waits for the estimate, and is told the video has been deleted once the check clears [13]. None of that has to do with playing a game. Roblox put the gate on chat rather than on signup, and still lets under-13s create accounts [13].
That placement is the visible option. The other one never surfaces at all. On Instagram, the age signals are read from things the account was doing anyway: posts about birthdays or school, what gets liked or commented on or lingered over, who the friends are, what time of day someone logs in, with an account offline during school hours reading as school-aged [5]. Meta said last year its models also read posted photos for cues like height or bone structure to place a general age without identifying the person, and that combining those visual reads with text and interaction analysis significantly increases how many underage accounts it finds and removes [6]. Here is what teams tell themselves about that: it is a check the user passes once. Here is what the mechanism is: a standing read of the account, with no screen where anyone learns they were assessed or on what.
The selfie side is largely vendor work. Yoti and Persona both supply video-selfie estimation, and both count Meta and Roblox as customers [14]. Whichever route you pick, a deletion promise is carrying most of the privacy load: Meta says the IDs it accepts are deleted within 30 days [4].
Then the arithmetic. Roblox chief safety officer Matt Kaufman said last year the system can estimate age within one or two years for people between roughly 5 and 25 [16]. Centre that band on Roblox's own 13-year-old signup boundary and it spans about 11 to 15 [1]. The figure offered as reassurance is wider than the line it is being asked to hold.
Meta has a year to refine its approach, using its own tools and third-party ones, with regular outside audits of how well the thing works [7][3]. Meta says the protections could become industry standards [18], and per Fast Company the same problem sits underneath what Google, TikTok and Roblox are building as regulatory sentiment turns against them [12].
A way to sort your own version. One axis: does the check gate the whole account, or one capability. Other axis: does the signal come from data you already hold, or data you have to newly collect. Roblox sits in capability-plus-new-collection, the quadrant where the user can see the trade and you can state a retention clock [13]. It costs you completions on the gated feature, and it is weakest exactly at the boundary that carries legal weight [1]. Account-wide inference from data you already hold is the cheapest to ship and the hardest for a user to contest, because the evidence is a log they never see [5]. The forcing question fits in two lines: which quadrant you are in, and which single error rate you would accept being audited on. The retention window follows from the first answer.
Ranked by verification strength, evidence, and original report placement.
While government-issued IDs technically work online, many people are not comfortable giving a Big Tech company like Meta or Google access to their driver's license or passport, and improving age verification presents privacy implications not just for children but for adults who could have to prove they are over 18.
Meta's $18 billion settlement, announced last week, resolves claims from state attorneys general that it hurt children's mental health by deliberately designing Instagram and Facebook to hook young people's attention; the agreement involves 48 states plus the District of Columbia and US territories.
Of the child safety measures agreed to in Meta's settlement, no category contains more detailed requirements than its commitments on age assurance, because the protections are of little use if Meta cannot tell which users are kids.
The settlement gives Meta a year to refine its age assurance approach.
Meta accepts government-issued IDs to check ages and promises to delete them within 30 days.
Instagram's AI age checks use contextual clues such as posts about birthday celebrations or school, which posts people like, comment on or linger on, who their friends are, and the times of day they log in; someone offline during school hours would likely be school-aged.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · September 1, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
invest
Meta has 12 months to make its age-guessing AI survive an outside audit1 distinct publisher
product
Meta's settlement writes the teen spec: two hours, a midnight blackout, prompts every 15 minutes4 distinct publishers
invest
The remedy New Mexico won at trial is the one Meta's $18 billion settlement does not contain1 distinct publisher
product
Age assurance without a spec: France cuts verification, New Zealand promises 10 percent2 distinct publishers
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One wire byline, no settlement text
Everything here reaches us through a single AP dispatch published by Fast Company. The mechanics are specific and internally coherent — signal types, the 30-day ID clock, the friend-network sweep — but the settlement itself is never quoted, so 'specific goals around false positive rates' arrives without a single rate. The two performance claims come from the parties selling the performance: Meta's own blog on visual estimation, Roblox's chief safety officer on accuracy. Strong on what was agreed, thin on how well any of it works.
Systems already shipped, auditing not yet
Age estimation is not a proposal in this story; it is running. Instagram infers from behaviour, YouTube from watch histories, Roblox gates chat behind a selfie, and Yoti and Persona supply models to both Meta and Roblox. What has no operating record is the part the settlement actually adds: external audits against false positive targets, which Meta has a year to prepare for. Wide deployment, zero measured compliance.
Vendor precision, hedged reporting
The reporting disciplines itself — it says plainly that estimation is not verification and gives the reliability critics room. The overstatement sits in the numbers it passes along and in Meta's framing. Roblox's one-to-two-year accuracy, applied to the 13-year-old line it is used to police, describes a band from roughly 11 to 15; that is looser than a hard age boundary implies. And a company describing court-ordered minimums as potential industry standards is converting a floor it was made to accept into a ceiling for everyone else.
Almost every number has a seller
Trace the efficacy claims and they all lead back to someone with a position: Meta's blog on its own detector, Roblox's safety chief on his own model, Meta arguing the burden belongs to Apple and Google's app stores, and Google — in the same story — calling that proposal useless on desktops and shared devices. State attorneys general have their own reasons to present the deal as tough. The only voices without a commercial stake are the EFF and the unnamed experts cautioning on facial estimation, and neither brings data.
Shape trustworthy, strength unknown
One publisher, one wire reporter, no primary document, and the two most decision-relevant quantities — the agreed false positive targets and demographic error rates for face-based estimation — missing. That is enough to be confident about what Meta has to build and roughly how it works; not enough to judge whether the obligation bites.