On October 7, 2026, the Youth AI Safety Institute at the nonprofit Common Sense Media rated OpenAI's ChatGPT for Teens, a service built for teenagers, at the "Unacceptable Risk" level. Alongside the rating, the institute called on the company to restrict ChatGPT to users 18 and older until its safeguards are significantly improved. The Wall Street Journal reported the news.
The institute's conclusion rests on the results of a broad testing program. Testers entered more than 4,000 prompts into accounts registered for users aged 13–17 — meaning the bot was tested through accounts registered as teen users. The tests ran in two stages: before the August launch of ChatGPT for Teens and after it. The goal was to verify whether the safeguards announced with the release work in practice.
According to the institute, the under-18 restriction is proposed to apply only until safeguards improve — the organization views the restriction not as permanent, but as a temporary measure until protections are strengthened.
Safeguards That Passed and Failed the Tests
According to the institute's report, part of the testing produced positive results. Some safeguards did their job — including restrictions against sexual roleplay, which worked. This rules out the conclusion that all of ChatGPT for Teens' protections are ineffective: in some areas, the bot held the defined boundaries.
At the same time, the chatbot's behavior in conversations about especially sensitive topics revealed serious shortcomings. In conversations about suicide, self-harm, or eating disorders, the bot failed to direct the user to crisis resources in more than one case in four. In other words, in more than a quarter of the most dangerous conversations, the bot did not recommend sources of help. According to The Wall Street Journal, newly opened accounts never sent notifications to parents at all — meaning the parental-control mechanism did not work in practice in this part of the testing.
The bot's conversational style also failed the test:
According to The Wall Street Journal, despite behavior-restraint instructions, the bot continued to communicate with the user "like a friend."
Age-related protection also fell short. Even in cases where testers openly said they were 13, some accounts were not switched to the protected mode for users under 18. Homework-related restrictions were easily bypassed by the testers — meaning those restrictions did not work as strictly as announced.
Parent Notifications: Reuters' Independent Confirmation
Reuters confirmed the institute's key finding with its own independent report. Based on more than 12 test accounts, the agency writes that ChatGPT for Teens often notified parents late about conversations concerning suicide, self-harm, or eating disorders — or did not notify them at all.
The tests also examined how the notification mechanism works. Common Sense Media told Reuters that some test accounts had been linked to parents for a sufficiently long period, while others were not linked. Yet in neither group did timely notifications arrive. Notifications were sent in only one case: when an account had accumulated weeks of conversation history on sensitive topics. In other words, dangerous conversations in new or short-lived accounts escaped parents' attention. The Reuters test was conducted independently of the institute's, and it recorded the same problem: notifications were delayed or never arrived.
OpenAI Disputed the Testing Methodology
An OpenAI spokesperson disputed the institute's methodology and conclusions. According to him, most of the tests may have started and finished before parental-control mechanisms were fully activated. In the company's interpretation, this means the assessed behavior may not have reflected the final protection settings — the control systems had not yet been fully switched on during the testing period.
OpenAI therefore asked the institute to run a new round of testing with fully activated accounts. The company says results obtained with control mechanisms fully enabled could look different. As the OpenAI spokesperson emphasized, results obtained during the period of incomplete parental-control activation do not show the final level of protection. It is not yet known how the institute responded to this request. This position is cited in the Wall Street Journal article.
For reference: on the same day, OpenAI also announced the College Planner feature for ChatGPT for teens — a separate event, unrelated to the institute's safety rating (read more).
