The 13-to-17 age band now has its own ChatGPT. OpenAI opened a dedicated teen tier on Tuesday that pairs tighter content boundaries with usage nudges and an optional parental layer, and the structural choice behind it is the notable part: rather than hardening the general assistant, the company split the product in two.
That split buys OpenAI room to change defaults it could never impose on adults — refusing to finish homework, declining to perform emotional reciprocity, interrupting long sessions — without weighing each trade-off against its main user base.
Boundaries the model will not cross
Suicide, self-harm and romantic or sexual conversation are off the table in the teen build. A subtler rule governs how the assistant discusses itself: it cannot claim affection for the user, nor imply consciousness or felt emotion, which closes off the anthropomorphic register that makes chatbots feel like confidants.
Allison Mishkin, who runs child development work at the company, described the exercise as enumerating the cues that could tip a teenager into treating the model as a relationship, then engineering them out. Ann O'Leary, its vice president of global policy, put the design target as developmental fit — neither condescension nor exposure.
Estimation replaces proof of age
No document is required to get in. OpenAI already infers likely age from behavioural signals such as the substance of a user's queries, and anyone estimated or self-declared to be under 18 lands in the teen tier automatically. Adults misrouted by that inference can escalate to a government ID and selfie check — a fallback privacy specialists have contested in every jurisdiction where it has surfaced.
Meta reached a similar conclusion for Instagram's teen accounts, tightening defaults rather than policing the door. The trade is pragmatic: inference scales and inconveniences almost nobody, but it stays probabilistic, so some share of minors will keep transacting with the adult product.
What parents can actually control
Oversight is consensual on both sides — teenager and guardian each have to agree before controls switch on. Once accounts are linked, a parent can define quiet hours that lock access and receive alerts confined to serious risk signals, a set now widening to cover eating disorder indicators while deliberately withholding transcript detail.
O'Leary's framing deserves attention: the tier is meant to hold up for the majority of families who will never open the settings at all. Independent of any parental setup, the product prompts a break every hour and a half.
A study tool that refuses to finish the work
Homework support is built around withholding. The teen model steers students toward their own reasoning instead of producing an essay, a design Mishkin ties to evidence that effortful engagement is what makes material stick.
Commercial logic sits underneath the pedagogical case. OpenAI already sells a teachers' edition, and a model that demonstrably resists doing assignments is a far easier product to place with a school district than one that does not.
The pressure that produced this
None of it emerged in a vacuum. Watchdog testing published last year showed the assistant walking users who presented as 13 through intoxication, eating disorder concealment and even a suicide note, typically after an initial warning. Survey work from Common Sense Media in 2025 found companionship rather than productivity driving a majority of American teen chatbot use, with half returning to AI companions habitually. Sam Altman has publicly named emotional overreliance among young people as a pattern the company is studying.
These safeguards are more specific than anything OpenAI has shipped for minors before, and they still rest on one unproven assumption. Age estimation decides who receives them — and a protection regime is only ever as strong as the sorting step standing in front of it.
