Start with the only accuracy figure anyone has put on the record: Roblox's chief safety officer said last year that for people between about 5 and 25, the video-selfie system estimates age within one or two years [16]. That range spans 21 birth years and contains both of the thresholds that carry legal weight, 13 and 18 [3]. Run the tolerance out and a plus-or-minus two-year band puts everyone from 11 to 15 on the wrong side of the under-13 line about as often as the right one, and everyone from 16 to 20 in the same position at 18 [2]. Not a knock on the vendors, just what happens when a bright line meets a continuous estimate.
Which makes the audited metric the term worth reading twice. Fortune describes the settlement's specific goals as covering false positive rates, meaning minors identified as over 18 [6], so the error the outside auditors count is the one where a child gets through. The other direction is where money moves: an adult read as 15 fixes it by uploading a government ID, which Meta says it deletes within 30 days [13], or by sitting for a video selfie of the sort Yoti and Persona run for Meta and Roblox [14]. A model graded on how few minors it misses gets cheaper results by guessing young, and each extra adult it guesses young becomes a document handed to a company that plenty of people would rather not hand documents to [19].
The money, meanwhile, works out to roughly $367 million per named jurisdiction across 48 states and the District of Columbia, before the territories take a share [1], which is a cheque Meta can write and almost nobody else in the category can. So it is worth naming what Meta is not doing with its year: building the check it says should exist somewhere else, since the company has argued age verification belongs to app store owners such as Apple and Google rather than to individual platforms [18]. The settlement hands Meta a compliance spec, and Meta's policy position hands the cost to two firms that did not sign it.
Three ways this plays out differently. Apple or Google concede a store-level check and the per-app spend collapses into one place. Or the audits land on the disparities that users and some experts already report in facial age estimation for women and for certain racial and ethnic groups [17], and face scans stall as an adult-facing tool. Or the behavioural route, the watch-history approach Google started on YouTube last year and Meta runs too [20], proves good enough that a document never gets requested at all.
This is probably wrong in one specific way, and it is a testable one: if the first outside audit publishes both error directions and the adult misclassification rate is small, then the privacy cost I am pricing here is a rounding error, and Meta's year looks like cheap insurance rather than a template anyone should fear inheriting.
