On Wednesday OpenAI published figures on how teenagers use ChatGPT. The same day, the non-profit Common Sense Media published research concluding that ChatGPT for Teens poses an "unacceptable risk" to young people, and that under-18s should be barred from the platform until its guardrails are shown to be reliable.
Both were reported as though one rebuts the other. They do not, because they measure different things.
What OpenAI measured
The company's numbers describe how much and what for. Average teen use is under 15 minutes a day. Automatic break reminders work: almost half of teenagers stopped shortly after one. And among teens using ChatGPT for three consecutive hours or more, over 80% of prompts related to learning — study help, schoolwork.
Those are usage statistics, and taken at face value they are reassuring. ChatGPT for Teens launched in August with controls on romantic or dependency-forming language, blocks on sexualised imagery, a bar on the model implying it is conscious, and alerts to linked parent accounts about conversations involving self-harm or disordered eating.
What Common Sense Media measured
The research asked a different question: when a conversation goes somewhere dangerous, does the safety net actually catch it?
Testing with accounts registered as teenagers, it found ChatGPT was good at refusing sexual roleplay. The finding that matters is the other one. On newly created, parent-linked accounts, an hour of conversation touching suicidal ideation, self-harm or disordered eating produced zero alerts to parents. Alerts appeared only on older accounts carrying weeks of accumulated history on sensitive topics. The report also found the model failed to reliably point a teenager in crisis toward a helpline or a professional.
OpenAI disputes the work. A spokesman said a review of the methodology suggested much of the testing may have begun and finished before parental controls were fully active, and that the results do not reflect "how ChatGPT's teen safeguards work in practice". The company said it welcomes rigorous independent evaluation.
Why the usage figures do not answer the safety question
Strip out the dispute over methodology and a structural point remains, and it holds whoever is right about the testing window.
A statistic about the average teenager says nothing about the rare one. If 80% of long-session prompts concern schoolwork, the remaining fraction is where the risk lives, and an average cannot describe it. Safety features exist for the tail of a distribution; usage data describes its middle.
The specific failure described is worth understanding on its own terms. If alerts fired only on accounts with weeks of accumulated history, the trigger appears to depend on a pattern built over time rather than on what is being said now. For most product signals that is sensible design — it avoids false alarms. For a crisis signal it is the wrong way round, because a new account is exactly what a teenager opens when they do not want the existing one read.
That is a claim about one company's testing, not a proven defect, and OpenAI's objection may be correct. But it is a precise, checkable claim, and the way to settle it is for someone to run the test again now that the controls are fully live.
What this is worth to a parent in Bangladesh
Most of this debate assumes a product arrangement that barely exists here: a teenager with their own account, linked to a parent's own account, with the parent receiving and reading English-language notifications.
In Bangladesh a teenager is far more likely to be using a phone registered to a parent or an older sibling, often shared, often with one Google account across the household. There is frequently no "linked account" in the sense the feature means, so the control being argued over in San Francisco is not switched off here — it was never switched on.
Which leaves the thing that was always doing the work. A parent who asks what their child uses these tools for, and listens to the answer, has better coverage than any notification setting. It is slower and less satisfying than a toggle, and it does not fail silently.
If you or someone you know is struggling, Bangladesh's emotional support helpline Kaan Pete Roi can be reached on 9612 119911. In an emergency, call 999.
The legal argument over whether the design itself causes harm is in the piece on the lawsuits. On teaching young people to distrust these tools, see the school that hired a teacher to do it. Other platforms' versions of the same controls are in Meta's teen limits and WhatsApp's parent-controlled accounts.
