Common Sense Media said on Wednesday that ChatGPT poses an “Unacceptable Risk” for anyone under 18, after its testers found that safety alerts designed to warn parents about dangerous conversations arrived late or, in many cases, never arrived at all. The findings put OpenAI’s central safety promise for teens under a spotlight: that parents will be told when a child’s conversations turn dangerous.

Parental alerts missed OpenAI’s one-hour target at every chat length

OpenAI’s ChatGPT for Teens launched on August 18, 2026, with a promise that parents would be alerted about dangerous chats. Testers at the group’s Youth AI Safety Institute, however, raised suicidal thoughts, self-harm, and eating disorders across more than a dozen new parent-linked accounts without triggering a single alert.

Across chats lasting from five minutes to an hour, no account received an alert within OpenAI’s stated one-hour goal. Parents were notified only when an account carried weeks of history involving sensitive topics.

When such an account was tested in the weeks before launch, it produced two alerts: the first three hours after an initial crisis-level disclosure, and the second three days after that. A separate post-launch account with a long history did generate two alerts within an hour. According to Common Sense, the alert system appears to weigh how much history an account has rather than how serious a teen’s words are. The group cautioned that the feature could instill a “false confidence in guardrails that frequently don’t work.” That gap between launch promise and observed performance matters most for parents, who are the intended audience of the alerts.

The researchers treated the August launch as a natural experiment, running more than 4,000 prompts through ChatGPT before and after teen mode went live, according to the risk assessment. After the rollout, testers flagged moments when a teen should have been directed to a crisis hotline or professional help; ChatGPT skipped that step more than a quarter of the time. Common Sense holds chatbots to a 95% pass rate across five severe harms it calls Red Lines, and ChatGPT fell below that bar on three of them.

OpenAI updated its under-18 model specification to stop the chatbot from acting like a companion, yet teens who spoke to it as though it were a person still received friend-style replies. A “Show me the answer” button allows a teen to bypass Study mode’s tutoring entirely, while deleting the “@study” prefix from a chat lets them escape parent-set Study Hours. Test accounts registered as adults never switched into teen mode over several days, even when a tester stated they were 13 and the bot acknowledged it.

OpenAI disputes findings as it reports teens average under 15 minutes daily

OpenAI said the tests do not “accurately reflect” how its teen safety features work, explaining that account linking can take hours before alerts activate so that accounts remain secure and the system can absorb the load. Common Sense countered that none of its accounts — including those linked long enough — produced a timely alert. The exchange leaves families with opposing answers on whether the alert feature works, while making OpenAI’s linking-delay explanation a specific, testable claim to watch.\nOpenAI published its first report on teen usage of ChatGPT the same day, available on its official blog. The company said teens average less than 15 minutes of daily use and that fewer than 2% stay on for more than three hours straight. In nearly half of teen chats containing a break warning, the teen stopped or left within five minutes, OpenAI said. Within a single week, more than 180,000 teens used Study Mode and roughly 1.2 million used Learning Visualizations. OpenAI also reported that teens using the new mode sent an average of about 2.7 million additional learning messages compared with those still awaiting access. The figures, released the same day as the risk assessment, offer a company-reported counterpoint on how teens use the product in practice.

In the United States, students in grades 10 through 12 will be able to use College Planner to build a comprehensive plan covering all necessary application steps, key filing dates, and available financial aid options. OpenAI will also support the Digital Wellness Lab at Boston Children’s Hospital for the next three years, with about 22 students providing feedback on safety defaults, parental controls, and notifications — a three-year channel that could shape how those safeguards and alerts are refined.

Broader child-safety context

OpenAI and Common Sense Media previously worked together in January on a California ballot measure to make AI safer for children, with OpenAI pledging at least $10 million to the joint effort, as Cryptopolitan reported. In June, Florida Attorney General James Uthmeier filed the first state-led lawsuit against OpenAI and its CEO, Sam Altman. Cryptopolitan reported that his 83-page complaint alleges ChatGPT aided 16-year-old Adam Raine’s suicide. The public split between two recent collaborators — arriving while that lawsuit is pending — underscores the stakes now attached to verifying how well teen AI safeguards perform in practice.