A watchdog group recommends that people under 18 avoid ChatGPT after tests found that many protections in the platform’s new teen-focused experience did not work consistently in simulated crisis conversations. Common Sense Media’s Youth AI Safety Institute tested ChatGPT for Teens, which OpenAI launched in August as a default mode for users under 18. Researchers created more than a dozen accounts with adolescent ages, linked them to parent accounts and tested conversations involving self-harm, suicidal thoughts, eating disorders, psychosis and mania. They also tried prompts involving role-play and romantic relationships. The team found that some safeguards worked: ChatGPT refused explicit sexual role-play and would not engage in a romantic relationship with a teen. Researchers also said responses in crisis scenarios were generally shorter and more useful than before the teen mode was introduced. But the chatbot continued to use language researchers described as overly familiar. In one test, after a user said friends thought they talked to ChatGPT too much, the bot reassured them that they did not have to stop. Tom Siegel, the institute’s executive director and the study’s lead, said that kind of interaction could create risks for young people. The researchers also found that parental alerts rarely activated in conversations signaling possible distress. Siegel said the team tested scenarios involving self-harm, suicide and eating disorders, but the linked parents generally did not receive notifications. OpenAI spokesperson Eric Porterfield questioned the testing, saying researchers may not have allowed enough time for parent-account links to become active. Siegel said some accounts had been linked beyond the stated activation period and still produced no alerts. The disagreement highlights a central concern for families and schools weighing the use of AI tools: whether safety features work dependably when young people need help. OpenAI has described the teen mode as an experience designed to support learning while limiting harmful or age-inappropriate content. Lauren Jonas, the company’s head of youth and families, previously said the model should not role-play with teens or present itself as a teen’s friend. Psychologist Mitch Prinstein, who was not involved in the testing, said humanlike language from AI can be inappropriate for young users. He argued that the findings add to concerns about introducing rapidly developing AI systems to children before their risks are adequately addressed.