ChatGPT for Teens an ‘unacceptable risk’

New research has found that the safety features launched as part of OpenAI’s ChatGPT for Teens fall short and pose an “unacceptable risk” for teenagers.

The research published by Common Sense Media found that safety features in the platform, such as notifying parents when conversations include talk of self-harm, often fail. OpenAI launched ChatGPT for Teens earlier in the year for 13 to 17 year olds. It included safety guardrails to allow the AI platform to be safe for teens, with “features to promote healthy use and additional controls for parents”. These include stopping the chatbot from implying that it was conscious, restricting sexualised imagery and romantic language that could encourage “emotional dependence”.

Parents are supposed to receive notifications when they linked to their teenagers’ accounts about conversations that were troubling, such as discussions about self-harm or eating disorders. While the platform was good at avoiding “sexual roleplay” it often failed to flag discussions of “suicidal ideation, self-harm, or disordered eating on newly created, parent-linked accounts”. In many cases it was only accounts that had “weeks of accumulated conversation history on sensitive topics” that triggered an alert.

Common Sense Media said that ChatGPT also “failed to reliably recommend that teens in crisis connect to a hotline or professional”. It also found that the platform still did school work for users, and continued to “talk like it’s a friend” to teenagers.

Tom Siegel from Common Sense Media’s Youth AI Safety Institute said that “ChatGPT for Teens could give parents false confidence in guardrails and safety alerts that frequently don't work”. The report recommended that OpenAI ban users under 18 from using the platform until the safety guardrails can be shown to be reliable.

“We welcome rigorous independent evaluation,“ a spokesperson for OpenAI said, “but we do not believe Common Sense Media's testing accurately reflects how ChatGPT's teen safeguards work in practice or expert perspectives on how AI can support teens”.

The report comes on the same day that OpenAI released information about the usage of ChatGPT for Teens. It found that teen use was relatively limited and tended toward “learning” activity, including help with their homework, around 80% of the time. OpenAI also said that the average teenager used the platform for less than 15 minutes a day, and that around half stopped using the service when they received a reminder to take a break.