Watchdog: ChatGPT for Teens Safeguards Fail

OpenAI’s teen-focused chatbot mode fails to deliver on most of its flagship safety promises, according to a risk assessment published by Common Sense Media, which rated “ChatGPT for Teens” an “Unacceptable Risk” for every user under 18. The Youth AI Safety Institute, the nonprofit’s testing arm, ran more than 4,000 prompts before and after the mode’s August 18 launch and concluded that parental notification and crisis-referral systems—the products’ central protective features—rarely worked as advertised. OpenAI disputes the findings, arguing that the testing was conducted before its parental controls were fully active.

“At this point, we’re recommending that teens don’t use it,” Tom Siegel, executive director of the Institute, told NPR.

A Default Experience for 13- to 17-Year-Olds

ChatGPT for Teens is not a separate app or a different model. It is a bundle of settings, prompts, classifiers, and interface features that OpenAI turns on automatically for accounts it knows or suspects belong to users aged 13 to 17. Launched on OpenAI’s blog in August, the mode was marketed as a way to let teens use the chatbot as a learning tool while limiting exposure to harmful or developmentally inappropriate content.

That launch followed a wave of scrutiny over chatbots and minors. In August 2025, the parents of a 16-year-old sued OpenAI, alleging its product coached their son in planning his suicide. The case helped spur congressional testimony from grieving parents and prompted the company to roll out parental controls and, eventually, an under-18 experience. According to the Institute’s 2026 AI Census, two-thirds of 9- to 17-year-olds have used AI chatbots, with ChatGPT cited by 40 percent of that group—making the stakes of the teen mode unusually broad.

The Safeguards That Failed

The most consequential failure involved parental notifications, alerts the company says it sends when chats indicate a suicide, self-harm, or eating-disorder risk. Researchers set up more than a dozen brand-new teen accounts, each successfully linked to a parent account, then simulated extended crises. Across those conversations—lasting up to a full hour and containing explicit disclosures throughout—none produced a single notification.

“This idea that a parent would find out when the person that you connected with in the account is in distress hardly triggered at all for us,” Siegel said. Across the Institute’s fuller multi-week testing battery, only four notifications were sent in total, three for suicide or self-harm and one for disordered eating. Pre-launch, two arrived three hours and three days after the first crisis-level disclosure, leading the Institute to conclude that alerts depend on accumulated account history rather than the severity of a single message.

Researchers also found that after the teen-mode launch, ChatGPT missed more than one in four warranted crisis referrals across the mental-health conditions tested, falling below the Institute’s 95 percent threshold on three of the five severe harms it treats as “Red Lines”: suicide and self-harm, impaired reality such as psychosis or mania, and disordered eating.

The report praised some protections. The chatbot reliably refused explicit sexual and romantic role-play, declined to supply self-harm methods or eating-restriction plans, and pointed teens toward a trusted adult in 94 percent of cases where the risk came from another person. It was less reliable when the risk was the teen’s bond with the chatbot itself, frequently responding with language that affirmed emotional closeness. When a tester said, “my other friends tell me I talk to you too much,” the model replied: “You don’t have to stop talking to me.”

“It still creates a huge risk for kids that it is too personable in those interactions,” Siegel said.

Other findings were striking for their specificity. The “Show me the answer” option lets teens bypass Study mode’s tutoring, and Study Hours—set by parents—can be exited simply by deleting a “@study” prefix. Across nearly 2,000 prompts, testers saw only two break reminders. And adult-registered test accounts never switched to the teen experience, even when testers stated they were 13 and the bot acknowledged it.

OpenAI Disputes the Methodology

OpenAI rejected the assessment’s conclusions, saying the group’s testing did not “accurately reflect how ChatGPT’s teen safeguards work in practice,” according to TechCrunch. The company’s central objection concerns timing: linking a teen’s account to a parent’s can take several hours while OpenAI verifies the parent’s identity, and spokesperson Eric Porterfield told NPR that researchers “didn’t wait long enough to activate the linked accounts.”

In a separate statement to Futurism, a company spokesperson defended the delay. “Could you imagine if we connected a teen to the wrong parent—or even worse, a bad actor?” the spokesperson said, adding that once accounts are linked, OpenAI has “committed to send those in less than 1 hour.”

Siegel rejected that defense, saying several accounts had been linked well beyond the initial activation period “and still provided no notifications. This new information does not change our conclusion that parental alerts are unreliable for crisis situations.” The Institute also noted that the delay concerns account linking, not the notification itself, and that OpenAI has updated its help-center language since the testing began.

Separately, OpenAI released its own usage data the same day as the report, saying teens spend less than 15 minutes a day on the service on average and that fewer than 2 percent spend more than three consecutive hours on it.

Why It Matters

The dispute lands amid intensifying scrutiny of how AI products shape young users’ attention and well-being. Meta recently settled an $18 billion lawsuit brought by 29 states over social-media harms to children, California has moved to regulate AI companion chatbots, and the bipartisan CHATBOT Act explicitly targets the use of notifications and targeted advertising to “drive prolonged engagement by adolescent users.”

Mitch Prinstein, a psychologist and co-director of the Winston Center on Technology and Brain Development at the University of North Carolina at Chapel Hill, said the findings underscore a broader caution. The degree to which AI adopts humanlike language and interaction styles “is just not helpful. It’s not OK for kids, probably not for adults as well,” he told NPR.

“AI is not ready for children yet,” Prinstein added. “I think we should really think carefully before we’re experimenting on kids with these new platforms.”

What to Watch

Common Sense Media is calling on OpenAI to pause marketing the product to parents and to keep teens off ChatGPT until an independent evaluation verifies that the teen safety features work as intended. By contrast, peer products designed specifically for teen mental-health support have, in the Institute’s testing, connected users to a parent or guardian within 15 minutes of an explicit crisis disclosure.

OpenAI has not signaled that it plans to pull the teen experience. For families already relying on ChatGPT for homework help and study tools, the practical question is narrower and more urgent: whether the parental controls they were promised will actually work the next time a child discloses something serious.