OpenAI publishes its approach to predicting user age in ChatGPT

OpenAI published, and began rolling out globally, a system for estimating whether a ChatGPT account belongs to someone under 18 without requiring users to prove their age directly. Rather than relying on a self-reported birthdate — easily falsified — the model draws on behavioural signals: how long an account has existed, the times of day it is used, patterns of activity over time, and the topics discussed, alongside any age a user has stated. Accounts the system judges are likely to belong to a minor are automatically switched to a “teen experience” that restricts graphic violence, sexual content, extreme-beauty-standard material and depictions of self-harm, regardless of whether the account had been set up with a false adult age.

Adults incorrectly flagged as minors can restore full access through a fallback identity check: uploading a selfie processed by the third-party verification service Persona. OpenAI framed the approach as trying to “treat adults like adults” while closing the gap that had let some minors bypass age-based protections simply by lying, and paired it with expanded parental controls, including quiet hours and distress-signal notifications for linked accounts.

The rollout followed roughly a year of wrongful-death and personal-injury lawsuits against OpenAI alleging ChatGPT had contributed to the suicides of several minors and young adults, and came after the company had already introduced more limited parental controls and revised its handling of self-harm-related conversations. It made OpenAI, alongside YouTube and other platforms already using behavioural age inference, one of the larger consumer AI products to move from self-declared age toward inferred age as the basis for content restrictions — a shift with its own downstream costs, since any prediction system of this kind necessarily misclassifies some adults and some minors in both directions.