OpenAI Started Guessing How Old You Are Yesterday. It Hasn't Said How Often It's Right.
· CX Pulse
OpenAI now predicts whether a ChatGPT account belongs to someone under 18 and switches the product automatically, without asking.
OpenAI launched ChatGPT for Teens on August 18. Most of the coverage went to the feature list, and it is a sensible list: no romantic language, no suggesting the model has feelings, a Study Mode that asks guiding questions instead of handing over the answer, homework reminders that trigger when someone looks like they are trying to shortcut an assignment, and safety alerts that reach a linked parent account.
The mechanism underneath the feature list got much less attention. That is the part worth reading twice.
You do not opt in
OpenAI's Lauren Jonas described how an account lands in the teen experience: "There's no need for a teen to create a new account or change anything; if we predict you're under 18, or you've told us so, this becomes your default experience."
Predict. There's no ID upload and no form to fill out.
The system reads signals already attached to the account. The general topics someone talks about, the times of day they use it, how and when the account gets used, and how long the account has existed. From those signals it estimates whether the person on the other end is under 18. If the estimate says yes, the product changes underneath them. It's rolling out globally to eligible accounts on both free and paid personal plans, with the rollout expected to finish inside two weeks.
OpenAI has not published how accurately that prediction performed in testing.
What this means for you
Take the subject matter away and what is left is software that looked at someone's behavior, reached a conclusion about who they are, and changed what they receive. Nobody was asked and nobody was told.
Your business almost certainly runs a version of this already. Fraud scoring at checkout. Risk tiers inside your payment processor. Spam filters on inbound email. Churn flags that quietly route an account to a different playbook. Chat routing that reads intent on the first message and decides which queue somebody lands in.
Those inferences are right most of the time, and when they're right nobody notices they happened. The entire cost of the system sits in the small share of cases where it's wrong.
Three questions are worth answering about every automatic classification running in your stack:
- Does the customer know a decision was made about them?
- If that decision was wrong, is there a way for them to say so?
- Does a person ever read what they said?
Most companies can answer the first one. A lot of them stall on the second. Very few have an honest answer to the third.
The imbalance is the useful part
What makes this launch a good example isn't the policy, which is defensible. It's the gap between the two halves of what got published.
The guardrails are documented in real detail, down to the kind of language the model won't use. The accuracy of the gate that decides who gets those guardrails is not documented at all. The rules are specific and the thing enforcing them is, publicly, unmeasured.
An adult who gets read as a teenager will open ChatGPT and find a product that won't talk to them the way it did last week, with no notification that anything moved. In any other industry that's a support ticket, and it's a bad one, because the customer can't describe what went wrong. They only know the thing they were using got worse.
That's the ticket your team will never solve quickly, and it is generated by a system nobody on your side thinks of as customer-facing.
The narrow lesson
Nobody is arguing teenagers should be handed an unrestricted chatbot, and a company with hundreds of millions of users has constraints a twenty-person business doesn't.
The transferable part is smaller than the headline. If something in your stack is allowed to decide something about a customer, then say how often it gets it right, tell the customer when it fires, and give them somewhere to go when it misses. The first one takes a query. The second takes a sentence of copy. The third is the one that costs real money, which is exactly why the first two get skipped along with it.
Worth an hour this week: list every place in your business where software sorts a customer into a category without being asked to. Then put a second column next to it: what a customer does when that category is wrong. The blank rows in the second column are the work.
Source: The Next Web, August 18, 2026