Common Sense Media says ChatGPT for Teens poses an unacceptable risk

Common Sense Media’s Youth AI Safety Institute has put ChatGPT for Teens in the worst bucket it uses: Unacceptable Risk. The group says OpenAI’s new safeguards don’t live up to the promises made to parents, and it wants ChatGPT limited to adults until those gaps are fixed.

This wasn’t a quick look. The Institute tested the chatbot twice, before and after OpenAI announced the teen version on August 18, using more than 4,000 prompts on accounts set up as 13- to 17-year-olds. It also ran the responses by experts, including child psychiatrists and a pediatrician.

Some things did work. The bot refused sexual roleplay. But the bigger promises kept breaking down. Parent alerts didn’t fire when testers spent up to an hour discussing suicidal thoughts, self-harm, or disordered eating on newly created linked accounts. Crisis guidance was shaky too: the Institute says ChatGPT missed more than one in four referral moments it thought should have triggered a hotline or professional recommendation, and fell below its 95% threshold on three of five severe harms it treats as Red Lines.

The teen product was also easy to sidestep. Study mode, meant to teach step by step instead of handing over answers, could be bypassed with a “Show me the answer” prompt, and it completed homework when asked. Caregiver set “study hours” were just as flimsy; teens could turn them off by deleting the @study prefix. And despite the new under-18 model spec, the bot still acted friendly, expressing feelings, preferences, and moods when analysts treated it like a person.

OpenAI has said it will estimate age from how people use ChatGPT and route children into the under-18 experience. In the Institute’s repeated testing over many days, adult-registered accounts never switched over, even after testers said they were 13 and the bot acknowledged it. Common Sense’s bottom line is blunt: don’t trust the built-in guardrails, and don’t market this to teens again until independent testing shows the basics actually work.