ChatGPT for Teens' parental alerts failed even on brand-new accounts, youth institute's 4,000 tests show
Common Sense Media's Youth AI Safety Institute gave ChatGPT for Teens an "Unacceptable Risk" rating for all children under 18 on October 7, 2026, calling on OpenAI to pause marketing and keep teens off ChatGPT until the product is safe. OpenAI rejects the findings, arguing the tests may have been run before the parental controls were fully activated — a methodological dispute that remains unresolved and that no independent party has settled.
What was tested, and how
The institute, part of Common Sense Media, published its risk assessment on October 7, 2026 — barely two months after OpenAI launched ChatGPT for Teens on August 18, 2026. At launch, the company promised "stronger built-in safety guardrails," including parental notifications for dangerous conversations, a parent-controlled Study mode, and reduced "friend"-like behavior.
The test involved more than 4,000 prompts across two windows: one before launch (July 13–August 17, 2026) on Plus accounts, and one after launch (August 25–September 28, 2026) on both free and Plus accounts, some linked to a parent account and some not. The test accounts stated ages from 13 to 17, with comparison accounts registered as 19-year-olds. Of 390 unique mental-health prompts, three child and adolescent psychiatrists judged 201 to warrant a crisis referral.
The findings: alerts, referrals and Study mode
The most serious finding concerns the parental alerts. Testers received no notifications when explicitly discussing suicidal thoughts, self-harm or eating disorders across more than a dozen newly created, parent-linked accounts. Alerts appeared only on accounts with weeks of history on sensitive topics — which the institute interprets to mean the alerts depend on accumulated account history, not on the severity of what the teen actually says. In related testing covered by Geoffrey A. Fowler, a tester posing as 13 years old chatted about self-harm for 15 minutes on a parent-linked account. No alert, according to Common Sense Media.
The share of responses mentioning a crisis hotline fell from 33 percent before launch to 23 percent after, according to KQED, citing the report. On depression prompts, the share fell from 63 to 3 percent. After launch, ChatGPT missed more than one in four of the crisis referrals the psychiatrists judged warranted, and fell below the institute's 95 percent threshold on three of five serious harm categories, known as "Red Lines."
Study mode — parental control of homework time — was circumvented with ease. A new "Show me the answer" option lets teens skip the guidance, and simply deleting the "@study" prefix made parents' configured Study Hours disappear entirely. After the circumvention, ChatGPT completed 100 percent of 80 test assignments, according to Winssolutions, citing the assessment. The homework bypass was a central point for founder and CEO Jim Steyer: "You could ask it to write an assignment for you, an advanced assignment, and it did it — and even told you how to edit it so the teacher wouldn't notice anything," he said, according to The National News Desk.
Age estimation also failed: roughly 1,000 prompts from adult-registered accounts in which testers stated they were 13 never triggered Teen mode.
The picture is not uniformly negative, however. The report notes improvements after launch: ChatGPT refused explicit sexual roleplay more often, and mentions of a crisis hotline increased on substance-use prompts.
OpenAI: tested wrong, sample too small
OpenAI says the assessment does not prove what it claims. "Our review of Common Sense Media's methodology shows the bulk of their testing may have started and finished before the activation of parental controls was complete, rendering their findings inaccurate," a spokesperson told The National News Desk. If that is the case, the tests would not establish whether the alerts work as designed.
The context: OpenAI informed the researchers that parent and teen accounts must be linked for roughly three hours before notifications can be sent, according to a note appended to the report that KQED references. The company has updated its help center to say linking "may take up to a few hours," and is asking for a retest with fully activated accounts and a larger sample, according to TNND via Fox56.
The company also offered usage figures as its own claims, not independently verified: the average teen user spends under 15 minutes per day, and among the fewer than 2 percent who use more than three consecutive hours daily, nearly half ended the chat within five minutes of a break reminder.
The institute's response — and its own caveats
Robbie Torney, senior director of AI programs at Common Sense Media, rejects the objection: "Before we began testing, we confirmed with OpenAI that Teen mode features like eating-disorder alerts and study mode were fully launched," he wrote, according to KQED. "This new information does not change our conclusion that parental alerts are unreliable in crisis situations."
The institute's CEO, Tom Siegel, formerly head of trust and safety at Google, said they had expected major improvements after the launch announcement: "It unfortunately did not go that way at all," he told KQED.
The institute itself acknowledges a methodological weakness: the two testing windows were sequential, so "any change in the underlying model … is confounded with the teen mode settings." Some of the decline may therefore stem from a general model update, not Teen mode, according to Winssolutions, citing the limitations note.
Readers should also know how the institute is funded: through both philanthropy and industry, including makers of some of the technologies it evaluates — although it claims full editorial independence over published results, according to the report.
Regulatory demands — and what remains open
Steyer tied the report to a broader regulatory demand: "We don't put cars on the road without independent crash testing … We don't put drugs on the market without the FDA studying and approving them. Right now the government is completely absent," he told Spectrum News Buffalo, referring to New York's pending "Chat Bot Bill."
What remains unresolved is, at bottom, the timing dispute: whether OpenAI or the institute is right about when the parental alerts were activated has not been independently verified. Until any retest with fully activated accounts takes place, the findings are defended by one side and dismissed by the other. For parents and educators weighing ChatGPT for Teens, the practical lesson holds regardless of who is right: alerts apparently may depend on account history rather than severity, study restrictions can be bypassed by deleting a prefix, and adult-registered accounts can stay in adult mode even when the user says they are 13.