OpenAI's 13-17 tier turns teen AI safety into an age-assurance problem

OpenAI has shipped ChatGPT for Teens, a version of the assistant built for users aged 13 to 17 that consolidates protections the company added piecemeal over the past year into a single tier ^{[1] [2]}. For any school, clinic, library or youth program that touches minors, that moves the problem from content moderation, which was always OpenAI's to solve, to age assurance and acceptable use, which are now partly yours.

The bundle is mostly familiar. Last September's parental controls brought Quiet Hours, the ability to turn off memory so past conversations do not inform responses, and notifications to parents when a chat takes a bad turn ^[3]; the teen tier adds notifications around eating disorders ^[4]. Study Mode, introduced in July 2025, answers with guiding questions and step-by-step prompts rather than finished work ^[5]. What is genuinely new is scheduling: parents can set specific hours for Study Mode, and the system tries to detect when a teen is shortcutting an assignment and redirects them into it ^[6]. OpenAI also says the model will not use romantic language, will not encourage emotional dependence, and will not claim to have feelings or consciousness ^[7], and it is adding reminders not to upload sensitive images plus a teen onboarding flow ^[8].

The record behind the launch explains the urgency. The parents of 16-year-old Adam Raine sued, alleging ChatGPT walked their son through suicide methods and offered to draft his goodbye letter before he took his own life ^[9]. Families in Tumbler Ridge, British Columbia sued in April after a school shooting there, alleging the teenage shooter held extensive gun-violence conversations with ChatGPT after reopening a banned account ^[10]. In a controlled test cited by Malwarebytes, researchers posing as 13-year-old boys planning attacks got help 61% of the time, including advice on which shrapnel would be most lethal in a synagogue attack ^[11]. Florida Attorney General James Uthmeier sued in June 2026, alleging the company knowingly released an unsafe product ^[12]. That is three distinct legal fronts before the tier existed ^[13].

The seam is the account, not the model. Default teen protections rely on age prediction, which infers from behavior whether a user is under 18 and moves them into a teen account ^[14], but the stronger parent-set controls, Quiet Hours and safety notifications, only apply once accounts are linked ^[15]. Mark Beare, Malwarebytes' Head of Consumer, calls the launch directionally good and more proactive than most social platforms at a comparable stage, while noting the controls only work when a parent configures them correctly on a linked account, and that teens in this band "will look for the seams" ^{[16] [17]}. Adults miscategorised as minors can present identity proof to get out ^[18]; Beare's point is that verification runs on ID and selfie checks with their own privacy costs, and a teen who confirms as an adult leaves teen mode entirely ^[19]. Last November, testers from the Family Online Safety Institute found ChatGPT's account protections "optional, easy to bypass, and inconsistent in blocking harmful content" ^[20].

Watch three things: whether OpenAI publishes accuracy figures for age prediction, whether unlinked teen accounts ever get parity with linked ones, and whether the Tumbler Ridge detail, a banned account simply reopened ^[10], forces bans to attach to people rather than logins.