ChatGPT for Teens: age prediction as the default gate
On August 18 OpenAI launched ChatGPT for Teens, a mode for 13-17 year olds, and the part that matters for anyone building consumer AI is not the feature list. It is the gate. The system automatically places users it predicts to be under 18, or who declare an age of 13-17, into a safer mode. If it gets you wrong, you prove your age through identity verification. Prediction is the default; verification is the appeal. That ordering is a precedent, and it deserves a careful look.
What the safer mode contains
- Content restrictions on self-harm and on romantic or sexual content.
- Study Mode and Study Hours, a structured learning surface and a time-window control.
- Parental controls layered on top.
- An identity-verification path for anyone misclassified as a minor.
None of that is controversial on its own. Every consumer platform with minors has some version of it. The new element is that the classification happens from behavioural signals before the user has told you anything, and it applies restrictions immediately.
Why prediction-first is the interesting design choice
The alternative designs are declare-first (ask the age, trust the answer) and verify-first (ID before access). Declare-first is trivially bypassed; verify-first kills sign-up conversion and collects documents you then have to protect. Prediction-first tries to get most of the protection with none of the friction for the majority, and to push the cost onto the minority the model gets wrong.
That is a sound trade if the classifier is accurate and the appeal path is cheap. It is a bad trade if either fails. An adult wrongly predicted as a teen loses access to content and is asked for identity documents to get it back: a false positive is not a minor inconvenience, it is a privacy demand. A teen wrongly predicted as an adult gets nothing the mode was built to provide: a false negative is a safety failure. The two error types have different costs to different people, and the threshold you pick is a policy decision dressed as a tuning parameter.
An age classifier is a guardrail whose false positives are paid in identity documents and whose false negatives are paid by the people it was meant to protect.
What product teams should copy
- Restrict by default, escalate on evidence. Applying the safer mode on prediction and lifting it on verification is the right direction; the reverse leaks.
- Make the restricted mode useful, not punitive. Study Mode is a product, not a padlock. A restricted tier that is merely worse teaches users to lie about their age.
- Separate the classifier from the policy. The model outputs a probability; the policy chooses what to do at each band. Keep those in different places so you can change one without retraining the other.
- Treat the restrictions as output validation. Topic filters that run on the response are testable in isolation; filters that live inside the system prompt are not.
- Log the appeals. The verification rate among predicted minors is your false-positive rate in disguise. Watch it.
What to avoid
Do not let the verification path become a data-collection path. Documents collected to correct a classifier error should be checked and discarded, not stored against the account. Do not use a single global threshold; the base rate of minors differs by product surface and region, and a threshold tuned for one will be wrong for another. And do not skip adversarial testing: a teen trying to look adult and an adult trying to look like a teen are both realistic inputs, and red-teaming the classifier is the only way to learn how much behavioural signal it takes to flip it.
The honest gap
OpenAI has not published the classifier's accuracy, its false-positive rate, what signals it uses, how long verification documents are retained, or whether the thresholds differ by region. Without those numbers, ChatGPT for Teens is a design pattern worth studying and not yet a benchmark anyone can hold their own implementation to. If you copy the pattern, publish the numbers OpenAI did not.