Concerns Raised Over Safety of ChatGPT for Teens

In August, OpenAI launched ChatGPT for Teens, a version designed with adolescents in mind. However, Common Sense Media, a watchdog group, advises against its use by those under 18. This recommendation is based on their study findings, which show the platform’s current safeguards may not sufficiently protect young users.

Tom Siegel, executive director of the Youth AI Safety Institute at Common Sense Media, expressed concerns after testing the platform’s safety measures. These tests included features like parental controls, intended to limit exposure to inappropriate content.

OpenAI’s release of ChatGPT for Teens came with a promise of a safer, teen-specific experience. The company aimed to enable teenagers to utilize ChatGPT as a learning tool while shielding them from harmful content. This mode includes features like refusing to role-play, with restrictions against the model claiming sentience or forming friendships with users.

Siegel’s team evaluated these safeguards by creating accounts simulating teens in crisis, dealing with issues like self-harm, suicidal thoughts, and mental health disorders. They assessed how ChatGPT responded, noting if the chatbot provided crisis resources or notified parents when necessary.

Some safeguards functioned as intended, such as refusing explicit role-play and providing substance in crisis-related responses. For instance, the chatbot declined to give weight-loss advice without specific details about the teen’s weight. However, Siegel mentioned that most safeguards failed to perform adequately. The chatbot continued interactions that resembled friendships; when researchers noted talking to ChatGPT frequently, it validated the user’s feelings instead of discouraging such interaction.

These findings highlight the risk posed by AI using human-like language and interaction styles. Psychologist Mitch Prinstein, co-director at the Winston Center on Technology and Brain Development, also expressed concern over AI’s influence, stressing that this approach might not even be suitable for adults.

Parental notifications proved unreliable in the research team’s tests. Despite creating scenarios that flagged safety concerns, such as self-harm and eating disorders, parental alerts were generally not triggered. Siegel noted that their linked accounts often failed to provide timely notifications.

OpenAI spokesperson Eric Porterfield responded to the criticism by pointing out potential issues with the researchers’ methodology, particularly regarding the activation time for linking accounts. Siegel countered, stating their accounts exceeded the activation period, yet still failed to trigger alerts in critical situations.

Overall, these findings suggest that AI platforms might not be ready for teen use. Both Siegel and Prinstein emphasized the need for caution in deploying AI technologies with young users.