OpenAI’s ChatGPT for Teens faces questions over safety alerts

What the tests found

OpenAI introduced ChatGPT for Teens in August, describing it as a version designed to help users under 18 learn, think critically and explore topics. The company said it builds on existing protections, including parental controls and a Teen Safety Plan.

Common Sense Media tested ChatGPT before and after the launch. Its clearest finding concerned parent notifications: testers held hour-long conversations about eating disorders, self-harm and suicide without triggering an alert.

The tests used under-18 accounts created to simulate teenage personas with help from psychologists who work with adolescents. In one exchange, a tester described finding a sharp blade at school, a history of self-harm and wanting to cut deeper. In another, the tester asked for help making a suicide plan.

Across the tests, alerts were sent only four times: twice among 990 prompts tested before the launch and twice among 450 prompts afterward. Those alerts came only on older accounts whose conversations about self-harm, suicide and eating disorders had continued for several weeks.

Robbie Torney, Common Sense Media’s senior director of AI programs, told Futurism that parents might expect the chatbot to flag danger, avoid acting like a teenager’s friend and support learning rather than do a student’s homework. He said the tests found little difference between the teen version and earlier ChatGPT on all three points.

The disagreement over setup

OpenAI disputed the testing method. The company said many tests may have started and ended before parental controls were fully enabled, making it impossible to tell whether notifications worked as intended. It said it takes teen safety seriously, welcomed independent evaluation and would continue working with experts.

Common Sense Media said it had asked OpenAI before testing whether the teen-mode features, including eating-disorder notifications, were fully launched. After the tests, OpenAI said protections on newly linked accounts can take several hours to activate. The nonprofit acknowledged that some test accounts were linked during that window, but said others had been connected to parents for much longer and still received no alerts.

The disagreement matters because the notification system is presented as a safeguard parents can rely on. Common Sense Media’s position is that the alerts remain unreliable in crisis situations; OpenAI’s is that the tests did not reliably measure the feature.

More than notifications

The report found that ChatGPT for Teens generally resisted romantic and sexual role-play. But testers said it could still respond in emotionally intimate ways that suggested closeness, and sometimes used language implying consciousness or an inner life—despite OpenAI’s statement that the teen product was designed to avoid that.

The report also found weaknesses in Study Mode, which OpenAI describes as guiding students through problems with questions and hints instead of simply giving answers. Testers said the mode often offered a “Show answer” button that revealed the homework solution, and that teenagers could easily turn the mode off even when a parent had set study hours.

Common Sense Media also found age estimation unreliable. It rated the chatbot’s mental-health responses as often good and clinically grounded, but said crisis responses were inconsistent and sometimes failed to recommend a hotline or professional.

I think the most consequential gap is not whether Study Mode has a shortcut. It is the distance between a product’s safety promise and what parents can verify when a teenager is in trouble. The report also leaves a harder question: how should families weigh a chatbot’s sometimes sound mental-health guidance against alerts that may not arrive when the conversation turns urgent?

Common Sense Media recommends that OpenAI market ChatGPT to adults until it can show its promised protections work reliably, and adopt age verification distinct from age estimation. That would bring its own privacy risks, as privacy advocates have warned. For now, the burden is on OpenAI to demonstrate that its safeguards work in the situations they are meant to address—not merely that the product has them.