OpenAI Will Put You in ChatGPT for Teens Without Asking

SAN FRANCISCO, 19 AUG 2026 — OpenAI has launched a separate ChatGPT for users aged 13 to 17, and it will place people into it automatically when its age-prediction system judges them to be under 18. Nobody has to opt in, and nobody is asked first.

The restrictions themselves are uncontroversial. The mechanism for applying them is not.

What launched

The teen experience restricts content around suicide, self-harm and romantic or sexual exchanges, adds homework and study support designed to help a student work through a problem rather than produce the answer, and includes measures intended to stop the assistant behaving like a person.

Lauren Jonas, who leads youth and families at OpenAI, described the routing plainly: a teenager does not need a new account or to change anything, and if the system predicts they are under 18, or they have said so, that becomes the default experience.

Prediction is a different thing from verification

Age verification asks for a document. Age prediction infers from behaviour, and the two fail in opposite directions.

Verification is accurate but exclusionary; it locks out anyone without the right document, a substantial number of people in this region. Prediction is inclusive but approximate; nobody is turned away, but some percentage of users will always end up in the wrong experience.

Both error directions matter here. A teenager the system reads as an adult loses the protections the launch exists to provide. An adult the system reads as a teenager finds a product quietly narrowed, on the basis of an inference they did not see, cannot inspect and were not told about.

Adults misclassified as teens will generate the complaints, and the public material is thinnest on the details of how that works. What signals feed the prediction, how confident it has to be, and what an adult does to correct it are not established in the material we could read.

Automatic enrolment is the design choice

OpenAI could have asked. It could have prompted uncertain users to confirm their age, or defaulted to the adult experience with an escape hatch.

It chose to route silently and default toward protection — a defensible, and deliberate, product decision. Defaulting toward protection means accepting a rate of false positives — adults with a restricted product — as the price of catching more minors.

Given the regulatory pressure, that is the rational choice. It is the direction every platform under scrutiny is moving, and it is far easier to defend to a regulator than the reverse. It also means the population of users affected by a misprediction is larger than the population it is designed for, because adults outnumber teenagers on the service by a wide margin.

The regional context is not hypothetical

Southeast Asia is where this argument is furthest advanced, and it is running in the opposite direction — toward documents.

Malaysia began enforcing a ban on social media accounts for under-16s on 1 June, requiring platforms with substantial user bases to verify age against government records such as MyKad or a passport, with penalties reaching millions of ringgit. More than seventy civil society organisations objected that mandatory identity checks across many platforms create their own privacy and surveillance risks.

The two approaches are not compatible. A regulator that requires verification against a government record will not accept a behavioural estimate as compliance, and a company that has built prediction precisely to avoid collecting identity documents will not want to abandon it. Any platform operating in both regimes ends up running two systems, and the weaker one defines the actual protection.

We reported that Vietnam has brought a high-risk AI list into force and that Malaysia has issued guidelines on automated decision-making and profiling. Age prediction is automated decision-making that alters the service a person receives. This falls squarely under Malaysia’s guidelines, and no platform has yet explained how its system complies.

The question nobody has answered about the data

An age-prediction system is, mechanically, a profiling system. It observes how a person writes and behaves and draws a conclusion about a protected characteristic.

That raises questions the launch material does not address. Whether the inference is stored against the account, how long it is kept, whether it is used for anything other than routing, and whether a user can see it are all unanswered, and each has a different answer under different data-protection regimes.

The awkwardness is that the privacy-preserving choice at the front door — inferring rather than demanding a passport — creates a new record at the back. A company that verifies holds a document check. A company that predicts holds a continuously maintained judgement about a user's age derived from their behaviour, which is arguably the more sensitive artefact of the two.

What a parent or a school should take from this

The protections are real and worth having. A teenager routed into this experience gets meaningfully better handling of the subjects that matter most, and the study behaviour — working through a problem rather than answering it — is the right design for homework.

Do not assume the system works for every teenager. Prediction is probabilistic, users are motivated, and the failure is silent: nothing tells a parent their child was misclassified as an adult. A school planning around this should treat it as a reduction in risk rather than a control, and should keep whatever supervision it would have had anyway.

What we could not establish

How age prediction works. OpenAI has published material on the approach that we could not open, and the reporting we could read does not describe the signals used, the confidence threshold, or how the model was evaluated for accuracy across regions, languages and writing styles — the last of which matters, because a system trained largely on English-language behaviour will not classify equally well everywhere.

Also unestablished: whether a misclassified adult can appeal or verify their way out and what that requires; whether teenagers can be moved into the adult experience by a parent; what happens to conversation history when someone is reclassified; whether the prediction runs once or continuously; whether any of this is deferred in jurisdictions with conflicting rules; and whether the system's error rates will be published.

What to watch

Watch first for accuracy figures from OpenAI. A safety mechanism whose error rate is unknown cannot be evaluated by anyone outside, and an independent audit of misclassification by region and language would tell parents and regulators more than any feature list.

Then watch whether a regulator accepts prediction in place of verification. If one does, it becomes the industry template because it is cheaper and less invasive than collecting identity documents. If Malaysia's approach prevails instead, platforms will be running document checks in some markets and inference in others.

Finally, watch what happens to the adults. The first serious complaint from a professional whose account was silently restricted will define how this is perceived. The question "why am I in this mode" needs an answer before that complaint lands.