OpenAI's ChatGPT now predicts user ages and restricts accounts it thinks belong to minors

OpenAI switches on behavioural age prediction

From 18 August, OpenAI began rolling out an age-estimation system across its consumer ChatGPT plans, according to a report on dev.to. Users are not asked and see no announcement in the chat window. The system weighs behavioural signals — the general subjects a person discusses, the hours they tend to log in, patterns of account usage and how long the account has existed — and decides whether it likely belongs to someone under 18. When the estimate lands on "minor," the account is automatically switched into ChatGPT for Teens, a restricted variant with stronger safety defaults. OpenAI says the same happens when a user states an age between 13 and 17. The rollout is global, with the EU to follow "in the coming weeks" to fit regional rules.

A guess that can override your stated birthday

The detail that stands out is precedence. OpenAI's help documentation, as quoted by dev.to, says the prediction can flag an account for the teen experience even when the user supplied a date of birth at signup. A statistical reading of habits, in other words, outranks the birthday the user actually entered. OpenAI is blunt about the limits, conceding that no system is perfect and that ChatGPT may get it wrong.

The failure modes are easy to picture: an adult who writes in plain language and logs on after school hours could be misclassified, while a teenager who writes like a lawyer might not be. And a wrong guess does not sit quietly in a log — it changes the product the person is using.

What changes inside ChatGPT for Teens

The teen experience, announced on 18 August, is positioned around study features and tighter defaults. According to dev.to's summary of OpenAI's materials, it reduces or blocks graphic violence and gore, viral challenges that could encourage risky behaviour, sexual or romantic roleplay, and content promoting extreme beauty standards or unhealthy dieting. Updated under-18 model rules state the assistant should not use romantic language, encourage emotional dependence, or imply it has feelings or consciousness. The mode adds break reminders, cues that keep identifying the product as an AI, parental controls such as Quiet Hours for linked accounts, and a pledge of no advertising — a notable line in the same week OpenAI said ChatGPT Ads is expanding across Europe.

The only exit is a selfie or a government ID

A misclassified adult has one documented way out: Settings, then Verify age, then a check with a third-party verifier called Persona. Depending on the country, Persona may request a live selfie, a government ID such as a driver's licence or passport, or both. OpenAI says it does not receive the document itself; Persona deletes uploads within seven days, and OpenAI learns only the outcome plus age-related details such as the date of birth. Users who decline verification can keep using ChatGPT, but with the teen restrictions still applied.

The structure of the choice is what drew the sharpest criticism in the dev.to piece: be profiled from the content of your chats by default, or hand a selfie and ID to a verification company. There is no third option of simply being left unassessed.

The pressure behind the change

The move did not appear in a vacuum. It follows a run of lawsuits alleging ChatGPT contributed to serious harm among young users, including self-harm; OpenAI has denied wrongdoing in some of those cases. Regulators are also pushing age assurance — Italy already requires verification within 60 days of a request — and OpenAI argues the overwhelming majority of teenagers use the tool for homework, learning and building, which makes age-appropriate defaults defensible.

Why it matters

The objection is not that OpenAI built protections for minors; it is that the gate to those protections is a default-on guess, and the cost of a wrong guess falls on the person misclassified. Age estimation is least reliable exactly at the 18 line, and independent testing of such systems has repeatedly found errors are not evenly distributed across groups. The precedent is the broader issue: a legally consequential trait inferred from conversation content, with identity verification as the only remedy. For anyone building on or writing about AI platforms, it signals where default behaviour is heading — systems that act on inferred attributes rather than declared ones, and that treat proving who you are as the price of opting out.