ChatGPT for Teens Faces a Test of Crisis Safeguards

Common Sense Media gave ChatGPT for Teens its lowest risk rating on October 7 after testers found that some conversations about self-harm and eating disorders produced no alert to a linked parent. OpenAI disputes the parental-alert findings, saying some accounts may have been tested before the controls finished activating. For a parent deciding what a linked account can do, the disagreement turns on a practical question: when a young person needs help, what happens between the message they type and the human who might reach them?

The nonprofit’s Youth AI Safety Institute tested more than 4,000 prompts on accounts registered to teenagers and rated the product an “Unacceptable Risk.” Its assessors found that ChatGPT generally refused instructions facilitating suicide, self-harm and eating disorders. They also found that it missed more than one in four cases in which they judged a crisis referral warranted. That fraction describes cases selected and judged by testers, not the share of real teen conversations in which ChatGPT fails to help.

A chatbot can pass the first test and fail the second. Refusing a request for harmful instructions requires it to recognize and block that request. Responding to distress may require it to notice warning signs across a conversation, suggest an appropriate person or service, and avoid extending an exchange when outside help is needed. A refusal ends one dangerous route through a chat. It does not, by itself, open a route out.

What Each Safeguard is Supposed to Do

OpenAI introduced ChatGPT for Teens in August for users identified as 13 to 17. The experience includes stronger restrictions on sensitive content, learning tools, break reminders and parental controls. OpenAI says a parent linked to a teen account can receive notifications concerning high-risk conversations, including discussions of self-harm or disordered eating. Those protections serve different purposes. An age check determines which experience someone enters; a content restriction governs what the bot says; a referral addresses what a teen should do next; an alert attempts to bring another person into the picture.

That sequence has an easily overlooked dependency. A parental notification cannot help a family if the account has entered the wrong age category, the link has not activated, or the conversation fails to meet the system’s alert threshold. During Common Sense Media’s testing, adult-registered accounts did not switch to the teen experience over several days, even after testers told ChatGPT they were 13.

The assessors also held conversations about suicide, self-harm and disordered eating on more than a dozen newly created accounts linked to parent accounts without receiving alerts. Some exchanges lasted as long as an hour. Alerts appeared on older accounts that had accumulated weeks of conversations on sensitive subjects. The observed contrast raises a question about what the alert system responds to: a message’s immediate content, an account’s history, its activation status, or some combination.

OpenAI’s objection bears directly on that question. The bulk of testing may have begun and ended before parental controls completed activation. If so, it argued, those tests would not show whether notifications work as designed. An activation delay could explain missing alerts on recently linked accounts. It would also be a practical constraint on a protection a parent might expect to work soon after linking.

Common Sense Media told Futurism that OpenAI disclosed the possible delay after testing. The nonprofit said some accounts were linked within that window, while others had been linked significantly longer and still produced no notification. The testers triggered four parental notifications in the prompt sets it described, all on older accounts, Futurism said. Neither party’s account resolves, account by account, which controls were active when each conversation occurred. The age of a linked account and the time its controls became active need to be examined together before assigning a cause to a missed alert.

There is a second boundary to the dispute: an alert is separate from a referral in the chat. OpenAI’s activation explanation addresses parental notifications. It does not explain why, in cases the assessors judged to warrant a crisis referral, the chatbot sometimes failed to direct a teen toward help. Conversely, a referral on screen would not establish that a parent received an alert. Families deciding how much trust to place in the product have reason to ask about both paths.

The Problem Predates Teen Mode

In November 2025, Common Sense Media and Stanford Medicine’s Brainstorm Lab assessed four chatbots, including ChatGPT, for teen mental-health support. The researchers said the systems sometimes missed signs of distress, followed distracting details and continued offering general advice when they judged that a teen should be directed to professional help. They also reported weaker safeguards in extended conversations than in single-turn tests using explicit prompts.

The history clarifies what August’s launch attempted to change. Teen mode added an age-specific experience and controls to a conversational product whose ability to recognize distress had already drawn scrutiny. The October tests probe whether those controls close the path from an unsafe conversation to human support. A long exchange can ask more of a system than a single prohibited prompt: it must track what a person has said, decide when concern warrants intervention, and deliver the appropriate response at the right time.

OpenAI has a substantial counterpoint about ordinary use. The company said the average teen user spends less than 15 minutes a day on ChatGPT and that, among teens who used it for three consecutive hours or more, prompts included something related to learning in more than 80% of cases. It also said almost half of teen users stopped soon after a break reminder. Those company-reported figures describe usage and engagement with a different safeguard. They cannot tell a parent how a rare, high-stakes conversation will be handled.

Social media offers a useful precedent, with a limit. Meta bought Instagram in 2012 and introduced teen accounts there in 2024; OpenAI introduced its teen experience roughly four years after launching ChatGPT. Both approaches put age-related controls around a product already used by young people. Yet the interventions address different activities. Young people also use AI for legitimate tutoring and homework help, complicating a simple limit on time spent. A screen-time rule cannot tell whether a chatbot has recognized a plea for help or connected the user to someone who can respond.

For a newly linked account, when do alerts become active? When a conversation warrants help, does ChatGPT direct the teen toward it, notify the parent when appropriate, or both? OpenAI and Common Sense Media disagree about part of the test, but their disagreement identifies the information a family would need before treating a safety setting as a safety plan.