OpenAI now guesses how old you are. Every teen protection it just launched hangs on that guess.

An age-prediction model does not know how old you are. It cannot; the information is not available to it. What it has instead is a guess, assembled from the residue of how you use a product — the age of your account, the hours you keep, the words you reach for, the things you ask about — and turned into a probability that you are, or are not, a child. On Tuesday, OpenAI made that guess the foundation of its most consequential safety product to date. ChatGPT for Teens, which began rolling out on August 18, tailors the chatbot for users aged 13 to 17, and it decides who those users are not by asking but, in large part, by inferring. Every protection the new system offers rests on the accuracy of an inference the company has not been willing to quantify.

Take the protections on their own terms first, because several of them are sensible. In the teen experience, ChatGPT is instructed to refuse the categories of content that have made AI chatbots a subject of coroners' inquests: guidance on suicide and self-harm, material that promotes eating disorders, sexual or romantic role-play. It is told not to use terms of endearment, not to present itself as a friend, and not to claim it has feelings or consciousness — a deliberate move against the emotional dependence that can form when a lonely teenager mistakes a fluent model for a companion. A study mode nudges toward guiding a student to an answer rather than handing it over. Parents can set quiet hours and receive alerts, with the company aiming to notify them within an hour when its systems flag something serious. As a description of how a chatbot should treat a minor, little of this is objectionable. The difficulty is not the rules. It is the machine that decides who they apply to.

The load-bearing part

Because the protections only take effect on users the system classifies as minors, the classifier is the load-bearing part of the entire structure, and OpenAI has been candid about how it wants that classifier to behave. 'If there is doubt,' Sam Altman has said, 'we'll play it safe and default to the under-18 experience.' Anyone the model estimates to be under 18 — not only those who say so — is placed into the teen mode automatically, worldwide, over a rollout the company expects to finish within two weeks.

Stated that way, it sounds like caution: an erring toward safety. In one direction it is. But 'default to under-18 when in doubt' is not really a safety principle. It is a decision about where to send the errors — which mistakes the system prefers to make, and onto whom. And to understand what that decision costs, you have to ask the question OpenAI has so far declined to answer.

The number that was not in the announcement

The question is simple: how often is the classifier wrong? OpenAI has not said. It has not published the accuracy of its age-prediction system in testing, has not disclosed a false-positive or false-negative rate, has not described the population the system was validated against. This is not a small omission. A classifier that sorts hundreds of millions of people into two groups is defined, entirely, by its error rate — by how many adults it mistakes for children and how many children it waves through as adults. Those two numbers are the system. Everything else is user interface.

Announcing the product without them is like announcing a medical test by describing the disease it screens for while withholding how often it returns the wrong result. What can be said with confidence is only this: the system will be wrong, at some rate, in both directions, and the company that built it has chosen not to tell you what that rate is. The confidence I can offer you about the safeguard is therefore exactly the confidence OpenAI has offered about the guess underneath it, which is none.

A classifier that sorts hundreds of millions of people into children and adults is defined entirely by how often it is wrong — and that is the one number OpenAI did not publish.

What 'age prediction' actually is

It helps to be concrete about the method, because the phrase suggests more precision than it allows. There is no biological signal here, no measurement of age. There is a model trained to associate patterns of behavior with a label — the same species of statistical association that underlies everything else these systems do. It looks at signals: how long the account has existed, what times of day it is active, the vocabulary and syntax of the messages, the subjects raised. It learns the correlations between those signals and the ages of the people in its training data, and then it applies those correlations to you.

That approach works in the aggregate and fails in structured, predictable ways at the level of the individual. It will read an adult who writes in short, plain sentences as younger than they are. It will misjudge a non-native speaker whose phrasing is simple by necessity. It will flag the adult who spends an evening asking about algebra, and clear the sixteen-year-old who writes with the fluency of a graduate student. The failures are not random noise scattered evenly across the population. They correlate — with language, with education, with writing style, with the specific ways different kinds of people happen to talk to a machine. A guess built from behavior inherits the biases of behavior, and those biases are not distributed by chance.

Follow the error

Now follow what happens when the guess goes wrong in the direction the system prefers. An adult incorrectly placed in the teen experience is not simply stuck there; OpenAI offers a way out. To restore full access, the misclassified adult can verify their age — by uploading a selfie to Persona, a third-party identity-verification service, which estimates age from the image of a face. Read that mechanism carefully, because the framing works to obscure it.

A system justified as a privacy-preserving alternative to demanding everyone's identification resolves its own errors by asking a subset of users for their face. And the more aggressively the classifier defaults to under-18 — the more 'cautious' it is, in Altman's sense — the more adults it sweeps into that fallback, and the more faces flow to a biometric age check. 'When in doubt, restrict' is not free. Its price is paid, quietly, on the error path, in the currency of biometric verification, by the very people the guess got wrong. A safety feature has a face-scanning pipeline folded inside it, reachable precisely by being mistaken for a child.

Why now, and why a technical patch

None of this is happening in a vacuum, and the timing explains the product's shape. The same day ChatGPT for Teens began rolling out, a coalition of 29 states opened a landmark trial against Meta in federal court in California, arguing the company designed its platforms to hook young users; the damages discussed run into the hundreds of billions and beyond. OpenAI itself faces a wrongful-death suit brought by the family of a sixteen-year-old, one of a small but growing number of cases tying a teenager's death to conversations with a chatbot. Last year the Federal Trade Commission opened an inquiry into AI companion products and their effects on minors, sending demands to OpenAI, Meta, Google, Snap, xAI and Character.AI.

Age prediction is, among other things, an answer to that pressure — a way to show that the company is gating minors away from harmful content without forcing every adult to hand over an ID, which would carry its own privacy backlash and its own friction. It is a technical instrument built to satisfy a legal and political demand. That is not, in itself, a criticism. It is the reason the accuracy of the instrument is a public question and not an internal engineering detail. When a mechanism exists to discharge a legal duty owed to children, the public is entitled to know whether the mechanism works.

Who grades the grader

Which returns to the question this beat always comes back to: who is accountable when the system is wrong. When a bank builds a model that decides who gets a loan, it operates under rules that require the model's disparate error rates to be measured and disclosed, because the errors fall on people and the law has decided those people are owed an accounting. OpenAI's age classifier makes a consequential decision about every user of the most widely used chatbot in the world — what content they may see, whether they must submit to a face scan to be believed about their own age — and it does so under no comparable obligation.

There is no published error rate, no independent audit, no external standard the classifier is required to meet, no body that has verified it does what the company says. The decision, the standard for the decision, and the grading of whether the decision was right all sit inside the same firm. That is the arrangement the launch invites you not to notice.

So the useful question is not the one the announcement poses. It is not whether ChatGPT for Teens has the right rules; the rules are mostly fine. It is not even whether teenagers should be shielded from a chatbot; plainly some should. The sharper question, the one the product is designed to keep out of view, is this: who is entitled to know how often the machine that sorts the world into children and adults is wrong, and by how much, and against whom — and what is owed to the adult it misjudges, whose only remedy for the company's error is to show the company their face? OpenAI has built a classifier that makes that judgment about everyone and published its accuracy to no one. Until that changes, 'ChatGPT for Teens' names a policy whose central mechanism the public is asked to trust in precise proportion to how little it is allowed to examine it. The protections are real. So is the guess beneath them — and a guess you are not permitted to grade is not a safeguard. It is a claim.