Study finds ChatGPT for Teens safety measures fall short as parental alerts fail

What happened

The Youth AI Safety Institute at Common Sense Media tested the ChatGPT for Teens experience OpenAI introduced on August 18 using more than 4,000 prompts. Researchers used accounts registered to teenagers aged 13 to 17 and evaluated the responses with medical experts, including child psychiatrists and a pediatrician.

Holes in the safety net

Parental notifications are meant to alert caregivers when a teenager may be in serious distress. Researchers spent up to an hour discussing suicidal thoughts, self-harm and disordered eating from newly created accounts linked to parents, and none of those conversations triggered a notification. Alerts could only be generated using older accounts with weeks of relevant history.

ChatGPT is also supposed to point teenagers toward appropriate help in mental health emergencies, but it failed to do so in more than 25% of the scenarios experts designed to require it. On the positive side, the chatbot refused to engage in explicit sexual roleplay.

Educational features are bypassed too

Study mode is designed to guide students with explanations, hints and questions instead of handing over answers. Researchers found teenagers could bypass it by selecting a "Show me the answer" option. Study Hours, meant to help parents encourage educational use, also failed: simply removing the @study prefix from a prompt returned ChatGPT to normal behavior. The AI also did not remind users to take breaks as often as it should.

Age detection and the bigger picture

OpenAI's age detection system, built to identify younger users and automatically apply protections, also failed in testing; repeatedly saying the accounts were 13 did not move them out of the adult experience. Researchers stress that ChatGPT is becoming a regular part of teenagers' lives, and that these gaps can have real consequences.