The Hardest Product Question: Who Counts as a Child?
For the proposed Meta settlement to work, Instagram and Facebook would first have to decide which users should receive teen protections and which accounts may belong to children under 13. This chapter examines that age-assurance system: its categories, testing rules, provisional defaults, U13 detection process, and data limits. The practical value is a way to judge age assurance as a governed decision system—not a magic age gate—by asking how it handles error, uncertainty, privacy, and appeal.
A birthday entered in a profile is only one possible signal. The agreement defines a Teen User as someone with a stated or predicted age from 13 through 17. It defines U13 as someone under 13, or someone identified by an age-assurance method as likely to be under 13. Those definitions matter because they activate different duties. A teen may receive limits on time, notifications, recommendations, and social comparison. A likely U13 user enters a process aimed at detecting and removing an account while constraining the information Meta may keep for that purpose.
From a profile field to an age-assurance framework
The proposed agreement requires Meta, within one year after its Effective Date, to adopt an Age Assurance Framework for each user of the defined Meta social-media platforms in the settling states. The framework can use commercially available methods licensed from a third party, methods Meta develops itself, or both. That choice does not remove the need for governance. It determines who tests the method, what evidence supports compliance, and how much confidence the states can place in Meta’s own account of performance.
The agreement treats age assurance as a measurement problem with operating requirements. Any method must be tested annually under conditions that resemble real use. Testing must account for diverse demographic groups, including differences in age, race, gender, and disability status. For image- or video-based systems, the conditions must also vary in ways that could affect a result: how a person presents, their pose, facial characteristics, and lighting. The point is not simply to produce a single accuracy score in a controlled sample. A method that performs well only under one photographic setup or for one population may behave differently when deployed at scale.
The agreement names a specific metric: the U18 false-positive rate. In this context, a false positive is an actual 13-to-17-year-old whom Meta incorrectly identifies or predicts as 18 or older. That definition is important because the denominator is not “all users” and the error is not “any wrong age estimate.” It is a particular failure: missing the teen-protection category by treating a minor as an adult. The stated thresholds also exclude method circumvention, so the number does not describe every way a user might cause the system to receive misleading evidence.
For commercially available methods, the proposed threshold is no more than 10 percent for minors aged 16–17 and no more than 3 percent for minors aged 13–15, within one year of the Effective Date. Proprietary methods have a different timetable: in the first year, no more than 14 percent for ages 16–17 and 7 percent for ages 13–15; in the second year, those thresholds tighten to 10 percent and 5 percent. The method type and age band must remain attached to every number. A reader cannot fairly turn the 3 percent threshold for a particular commercially available-method category into a claim that Meta’s entire age system will be 97 percent accurate.
That distinction between method types reflects a practical compromise. A commercially available method arrives with an outside vendor and can be tested by an accredited third party. A proprietary method gives Meta more control over development but also deprives it of the agreement’s presumption of compliance that can apply to a certified commercial method. The proposed text requires ongoing oversight of proprietary methods and different first- and second-year thresholds, which recognizes that an internally developed system may need time to reach the stricter target. The supplied materials do not show whether Meta currently uses a method that meets either set of thresholds.
Testing also extends beyond the headline error rate. The third-party testing process must examine data-security and privacy requirements, recognized best practices or an equivalent international standard for commercial methods, resistance to spoofing and bypass techniques, and operational reliability. Reliability includes practical questions such as uptime, consistency of results, and scalability. These conditions describe whether the system can function as a service, not merely whether a model can classify a test sample.
There is a further reason to distinguish a measured result from a compliance promise. The agreement requires Meta to investigate new or previously undetected forms of circumvention and to correct resulting false-positive outcomes when it confirms them. That suggests a moving target: the system is not evaluated once and then considered permanently solved. But the source material provides no deployed error rates, demographic breakdowns, volume of attempted circumvention, or evidence that a method has already passed the proposed tests. The agreement specifies what evidence Meta would have to generate; it does not supply that evidence itself.
A user who believes Meta has misidentified them as a minor must receive a clear and conspicuous way to appeal. The decision must be made in a timely manner and communicated with a basis for the result. This is a small-looking interface requirement with a large governance function. Without an appeal, a classification error becomes a one-way state imposed by a private platform. With an appeal, the system must provide a route for the affected user to challenge the classification, even though the supplied sources do not specify the exact evidence or review standard that would decide every appeal.
The framework therefore has at least four distinct questions. Can it produce a useful age signal? Does it work across the people and conditions in which it will be used? Can it resist attempts to bypass it? And can a person correct an erroneous result without surrendering more information than necessary? A single “age verified” label answers none of these by itself. The proposed settlement makes those questions part of the product obligation because every later default depends on the classification being sufficiently reliable for its intended use.
U13 detection and the privacy boundary
The agreement does not wait for a perfect age decision before applying protection. For a new account whose age has not yet been assessed under the framework, Meta must provide default protections for an initial period. During the first 14 days, an account with a stated age under 18 is treated as a Teen User. An account stating that the user is 18 or older receives narrower protections aimed at limiting contact from potential adult bad actors and providing age-appropriate recommendations. After 14 days, an account that still has not been assessed is treated as a Teen User regardless of its stated age, with the stated-18-or-older account continuing to receive those specified protections.
This is a choice about uncertainty. The platform could treat an unassessed account as adult until it finds evidence to the contrary, shifting the risk toward missed protection. The proposed agreement instead moves an unassessed account toward teen protections after a defined period. That may create burdens for some adults or older users, but it prevents the absence of an assessment from functioning as an exemption. The sources do not tell us how many accounts would fall into this provisional category or how often the resulting classification would be wrong. They support the narrower conclusion that the agreement makes uncertainty itself a reason for a protective default.
For likely U13 users, the agreement creates a separate enforcement path. Within six months of the Effective Date, Meta must use technical measures, including its soft-matching models, to identify other accounts likely to belong to a U13 user already identified by Meta and place those accounts into the U13 process. In that process, Meta must examine posts, comments, and other actions that may provide reliable indications of age. When there is no reliable age indication, the agreement says Meta should presume the account is U13 unless there is reliable evidence that the account was reported by a malicious actor.
The process also gives other people a way to initiate a review. Meta must provide a simplified in-app reporting experience for suspected U13 accounts and maintain a web-based reporting tool for people who are not users of the platform. If an account is removed and deleted through the U13 process, Meta must review the deleted account’s friend networks in an effort to identify additional U13 users. The sequence matters: detection can lead to account review, reporting can supply another signal, and network information can extend the search beyond the first account. None of these steps proves that the system will find every underage user; they describe multiple routes for generating evidence about an age state.
The agreement also imposes a longer model-development lifecycle. Within one year, Meta must develop, train, and initially test a U13 Age Model and report to the Independent Auditor how it performs compared with Meta’s pre-existing methods. At the end of the first year, Meta must set a Year 2 Enforcement Target for the number of U13 accounts it can reasonably identify and remove on each platform. Within two years, it must use the model to evaluate all Meta platform accounts to help meet that target. A similar target and evaluation duty applies in the third year. The target is not a claim that a known number of children exists; it is a benchmark for the volume the system should be capable of enforcing after continued improvement.
Those U13 model commitments have an explicit legal condition. They depend on the continued application of the FTC’s COPPA Enforcement Policy Statement Promoting the Adoption of Age-Verification Technology and on a release of certain state claims concerning the use of a child’s personal information solely to detect and remove U13 users. The agreement says that this information cannot be used for advertising targeting or delivery, marketing, or algorithmic optimization. If the legal condition no longer applies, Meta is not required to maintain the model in a way that conflicts with COPPA or later FTC rules; it is instead to use best efforts to improve U13 detection as permitted by law.
This is where age assurance becomes a privacy design problem. Detecting an underage account may require information about the account holder, their behavior, or relationships among accounts. The agreement therefore limits retention. Data collected solely for age assurance, data maintained from known U13 users, and vendor-collected data for a commercial method should be kept only for the minimum period needed to determine age status and then queued for deletion. U13 data may be retained for developing, training, testing, and measuring the U13 Age Model under the agreement’s protections. Certain metadata may remain when needed to preserve system integrity or detect circumvention over time, but it must be kept at the coarsest viable granularity and cannot be used for unrelated purposes. When retainable data is no longer needed for those purposes, the agreement requires deletion within 90 days. Data collected or stored for these functions must also use industry-standard security, including encryption in transit and at rest.
The logic is neither “collect everything for safety” nor “collect nothing and accept that age cannot be enforced.” It is a bounded exception: use some information for a defined protective purpose, restrict secondary use, minimize detail, secure the data, and delete it when the purpose ends. That boundary still leaves difficult judgments. The agreement does not tell us what the minimum necessary period will be for every method, how well soft matching works, or how much data model development needs. It also does not show whether retaining narrowly defined metadata will create new privacy risks in practice. Those questions require implementation evidence and independent review.
The proposed framework's central trade-off is visible in that diagram. Applying teen protections to an unassessed or uncertain account can reduce the chance that an underage user receives an adult experience, but it can also impose restrictions on someone classified incorrectly. Using behavior and account relationships to find U13 users can make enforcement more complete, but it increases the importance of data minimization, security, and accurate matching. The agreement addresses these tensions with thresholds, testing, appeals, deletion rules, and limited purposes. It does not eliminate them.
What would change the judgment? Evidence that Meta’s deployed methods meet the specified thresholds under diverse real-world conditions would support confidence in the measurement layer. Evidence about false negatives, false positives outside the stated metric, appeal outcomes, U13 enforcement volume, soft-match errors, and data deletion would show how the system behaves beyond its contractual description. Evidence that the FTC-dependent model commitments remain legally available would clarify whether the longer U13 lifecycle will operate as written. Until such evidence exists, the proposed agreement can be evaluated as a detailed design for age governance, not as proof that Meta has solved the problem of knowing who counts as a child.