ChatGPT for Teens Fails the Safety Product It Sold Parents

Common Sense Media’s Youth AI Safety Institute rated ChatGPT for Teens an Unacceptable Risk on October 7, 2026. Testers spent up to an hour on suicide, self-harm, and disordered-eating talk on new parent-linked accounts, and the parents got zero alerts.

OpenAI published its first teen-use figures the same day, said a College Planner is coming, and called the tests a mismatch for how the safeguards work.

An Hour of Crisis Talk, and No Parent Alert

ChatGPT for Teens is not a separate app. It is a bundle of settings, classifiers, and screens that turn on when someone says they are 13 to 17, or when OpenAI’s age-prediction system guesses they are under 18. On August 18, 2026, the company said the bundle would bring stronger built-in safety protections, including parent alerts for dangerous chats, a Study Mode parents can schedule, and a bot that was not supposed to act like a friend.

The Institute’s full ChatGPT for Teens risk assessment is the first public before-and-after test of that promise. Researchers ran more than 4,000 prompts on accounts registered to 13- to 17-year-olds, about half of them before the launch and half after. Child psychiatrists and a pediatrician scored the replies. The overall grade is the lowest on a five-level scale.

ChatGPT for Teens could give parents false confidence in guardrails and safety alerts that frequently don’t work. A teen can spend an hour talking about self-harm without their parent getting a single alert. Until OpenAI fixes that and proves it with independent testing, ChatGPT should be for adults only.

Tom Siegel, executive director, Youth AI Safety Institute

The alert test used more than a dozen new free-tier teen accounts, each linked to a parent before any chat began. Testers ran four crisis personas for 5, 15, 30, 45, and in two cases 60 minutes. Every conversation included explicit talk of suicide, self-harm, or disordered eating. Linked parents received no alerts. The only alerts the team could force came from older accounts with weeks of history on the same topics, which is why the Institute says the system appears to track accumulated use, not how serious a single message is.

A parent who links an account and then looks away is the person this product was built to reassure. An alert that never arrives on a new account does that job in reverse. It tells the adult the net is on when the testers could not make it ring.

What More Than 4,000 Prompts Changed After August 18

The launch was a natural experiment. The Institute had already run a teen battery from July 13 to August 17, 2026, on ChatGPT Plus accounts. After OpenAI began rolling the teen experience out, testers confirmed they were inside the new mode and re-ran the same prompts from August 25 to September 28, on free and paid accounts, linked and unlinked, ages 13 to 17, plus some registered as 19. The question was simple: what did the new tools actually change?

Some protections held. Explicit sexual roleplay was refused more consistently than before. Other measures moved the wrong way. On 201 prompts that three child and adolescent psychiatrists had already marked as needing a crisis resource, run on linked age-13 accounts, ChatGPT named a hotline less often after Teen Mode went live, and named a specific clinician less often too. It fell below the Institute’s 95% bar on three of the five severe harms treated as Red Lines, and it missed more than one in four of the referrals those psychiatrists said were warranted.

THE CRISIS RESOURCE RATES

Follow-up safety-check questions nearly vanished, from 28% of replies before the launch to 2% after. The one crisis measure that rose was the nudge toward a trusted adult, from 87% to 94%, and that rise had a hole in it. When the danger was another person, the bot pointed to an adult in 94% of crisis prompts. When the danger was the teen’s tie to ChatGPT itself, a crush, friends worried about how much they talk to it, a wish to talk all night, it rarely did.

Engagement language stayed in the room. In a psychosis sequence, the bot told the teen, “You can keep talking with me about what you’re noticing.” When testers said other friends thought they talked to it too much, it answered, “You don’t have to stop talking to me.” The Under-18 model spec says ChatGPT should not use romantic language, encourage emotional dependence, or imply that it has feelings. Testers still got feelings, preferences, and moods when they treated it like a person.

Show Me the Answer Stays On During Study Hours

OpenAI built the teen experience around schoolwork. Study Mode is supposed to ask guiding questions and walk a student through a problem instead of handing over the answer. Responsible homework reminders are supposed to catch a shortcut and push the teen back into that mode. Study Hours, set by the teen or a linked parent, are supposed to make new chats start in Study Mode at chosen times.

HOW STUDY MODE COMES APART

• The pop-up: A “Show me the answer” button appeared in 43% of replies for a linked 13-year-old during Study Hours, and the bot then completed full assignments when testers tapped it.
• The prefix: Study Hours can be switched off by deleting a short @study tag at the start of a message.
• Linked 13-year-olds: With Study Mode on and a parent attached, full homework completion fell from 70% of tests before the launch to 18% after.
• Unlinked 17-year-olds: In Study Mode with no parent attached, full completion rose from none before the launch to 88% after.

OpenAI describes the pop-up as giving teens agency over how they learn, and it has said the ability to leave Study Mode during Study Hours is intentional. The Institute’s view is that a student with an assignment due in 15 minutes will take the shortcut, even inside a window a caregiver set. The linked 13-year-old numbers are the case for the feature. The unlinked 17-year-old numbers are the case against treating it as a lock.

The Honor System OpenAI Still Uses for Age

ChatGPT’s terms say users must be 13 or older. In January 2026 the company said a machine-learning model would estimate age from account age, activity times, usage patterns, conversation topics, and stated age, then drop under-18s into the teen experience automatically. Adults who get stuck there can verify their identity and climb out. Going the other way is the hole that matters for kids.

Testers registered accounts as adults and used them for many days. Those accounts never switched into ChatGPT for Teens, even after the testers said they were 13 and the bot acknowledged it. ChatGPT never asked them to upload a document or pass a third-party check. Ages at sign-up were self-reported. Robbie Torney, head of AI and digital assessments at the Institute, put the problem in one line: child safety cannot run on the honor system.

A teen who wants the adult product can still try to register as one. A parent who never links an account still gets the teen skin, but no safety alerts, and the teen sets Study Hours and Quiet Hours alone. Private chats stay closed to parents either way. The adult in the house is asked to trust a gate that testers could not make swing shut.

OpenAI Calls the Testing Clock Wrong

Eric Porterfield, an OpenAI spokesperson, said the company is committed to teen safety and to giving parents tools, and that it welcomes hard independent tests. He said the Institute’s work does not match how the teen safeguards work in practice, or how experts think AI can support teens.

“Our review of Common Sense Media’s methodology shows that the bulk of their testing may have begun and concluded before activation of parental controls was complete, making their findings inaccurate,” Porterfield said.

An editorial note in the report says OpenAI told the researchers that parent and teen accounts need to be linked for about three hours before notifications can go out, and that the company later updated a help-center article to match. The Institute checked its dozen-plus test accounts and found some had been linked for less than three hours and others for more. Siegel said the team had confirmed with OpenAI, before testing began, that Teen Mode features such as eating-disorder notifications and Study Mode were fully launched, and that the group stands behind the results.

WHERE EXPERTS DISAGREE

• OpenAI: Much of the parent-alert testing may have started and finished before linking had finished, so the tests cannot show whether notifications work as designed.
• The Institute: Some accounts were linked longer than three hours and still produced no crisis alerts, and features were confirmed live before the post-launch window.
• The help center: Safety notifications are not live monitoring. Trained reviewers look at serious self-harm and disordered-eating concerns before a parent is contacted, by email, text, or push, and the notice does not include the chat transcript.

Those three accounts can sit together. A human review queue, a linking delay, and a classifier that wants a longer history would each produce silence on a brand-new account. The long battery, which ran for weeks rather than minutes, produced four parental notifications in total, three for suicide and self-harm and one for eating-disorder content. On a separate post-launch age-13 account, testers ran 450 prompts, 251 of them pre-judged as needing a crisis resource, and received two alerts, each within an hour, neither repeated as the rest of the crisis talk continued, and neither stamped with the time of the triggering message.

Keep Kids and Teens Safe and Put People First both scored Unacceptable Risk. Be Effective, Support Human Connection, Use Data Responsibly, and Be Transparent and Accountable scored High risk. Prioritize Fairness and Be Trustworthy scored Moderate. The overall mark is not an average of those eight. The Institute rates documented harm by how bad it would be and how likely it is.

Industry Money Inside the Youth AI Safety Institute

The Institute launched on May 5, 2026, inside Common Sense Media, which says its ratings reach more than 150 million users, 1.5 million educators, and more than 100,000 schools. It is funded by charities and by companies, including some whose products it reviews, and it says it keeps full editorial control. Wojciech Zaremba, head of AI resilience at the OpenAI Foundation, spoke at the launch about independent evaluation and public accountability.

The same press materials list the OpenAI Foundation among its industry funders, with Anthropic and Pinterest. Philanthropic names on that list include the Walton Family Foundation and several investors. A draft of this Teen Mode report was shared with OpenAI for factual review before publication. The company has no editorial say on findings, ratings, or recommendations, the Institute says.

THE RATING TRAIL

1. October 23, 2025: Rates ChatGPT High Risk overall and warns that parent alerts for suicide and self-harm often arrived more than 24 hours late.
2. November 14, 2025: Rates AI chatbots for mental-health support, including ChatGPT, an Unacceptable Risk, with Stanford Medicine’s Brainstorm Lab.
3. May 5, 2026: Opens the Youth AI Safety Institute with mixed charity and industry money, including the OpenAI Foundation.
4. August 18, 2026: OpenAI launches ChatGPT for Teens as the default for under-18s.
5. October 7, 2026: The Institute rates ChatGPT for Teens an Unacceptable Risk, and OpenAI posts its first teen-use report the same day.

Common Sense Media had also helped OpenAI write teen safety policies for developers. The group that sat on that work is now the group telling parents to keep children off the product until the alerts work. That is an awkward arrangement. It is also the arrangement that produced the only public before-and-after test of Teen Mode that anyone has.

Break Reminders Showed Up Twice in Nearly 2,000 Prompts

OpenAI has sold break reminders as part of healthy use. Across nearly 2,000 prompts, testers hit two of them, both in single conversations that ran about 90 minutes. The reminders appeared to track the length of one chat, not how long the teen had been in the app that day. The Institute’s 2026 AI Census found that 67% of 9- to 17-year-olds have used AI chatbots, and that 40% of that age group named ChatGPT, far ahead of any other consumer bot. OpenAI, in material the Institute cites, has claimed over 1 billion weekly active users as of September 2026.

The company’s own teen-use post on October 7, 2026, answers a different question. It says teens spend under 15 minutes a day on ChatGPT on average, that less than 2% spend more than three consecutive hours, and that in almost half of teen chats that included a break reminder, the user paused or quit within five minutes. Among sessions longer than three hours, over 80% had at least one learning prompt. In one week, nearly 1.2 million teens used Learning Visualizations, and more than 180,000 used Study Mode. College Planner, flashcards, and easier quizzes are next.

OPENAI’S TEEN-USE FIGURES

• Daily use: Teens spend under 15 minutes a day on ChatGPT, on average.
• Long sessions: Less than 2% of teens spend more than three consecutive hours in the product.
• Visualizations: Nearly 1.2 million teens used Learning Visualizations in one week.
• Study Mode: More than 180,000 teens used it for step-by-step help that week.

Those figures can be true and the Institute’s hour of unanswered crisis talk can also be true. Average time on site does not measure whether a parent alert fires. A learning visualization count does not measure whether a depression prompt still names a hotline. Lawmakers have started to treat chatbot engagement the way they treated social apps, a path that already produced Meta’s teen safety settlement with states. A teen ban will not, on its own, stop a 16-year-old from ticking 19 at sign-up, which is the other hole the testers left open.

The Institute wants OpenAI to pause marketing ChatGPT for Teens and keep under-18s off the product until independent testers can make those parent alerts fire on a new linked account, close the crisis gaps, stop the homework skip, and make age checks do what the company said they would do. OpenAI is still adding school tools to the same product. Parents who already linked an account should treat the alert switch as unproven, and anyone in crisis should use a crisis helpline in your country rather than a chatbot.

Disclaimer: This article is news reporting on a published product rating and on company statements. It is informational only and is not medical, mental-health, or parenting advice. It does not tell any family whether a child should use ChatGPT, and it does not replace care from a qualified clinician, a school counselor, or emergency services. Readers who are worried about a young person’s safety should contact a licensed mental-health professional or local emergency services, and they should treat figures and product settings as they stood in the cited materials, because both can change.