New ChatGPT mode for teens criticized due to safety system failures

Common Sense Media has called the ChatGPT for Teens mode launched by OpenAI in August an "unacceptable risk" for users under 18. During independent testing, experts found problems with parental notifications, crisis response, and educational restrictions.

The study was conducted by the Youth AI Safety Institute, part of Common Sense Media. Specialists tested over 4,000 queries on accounts registered to users aged 13 to 17. About half of the tests were performed before the launch of ChatGPT for Teens, and the rest after. Responses on mental health topics were additionally evaluated by child and adolescent psychiatrists and a pediatrician.

One of the main complaints was parental notifications about potentially dangerous situations. Researchers created over a dozen new teen accounts and linked them to parent accounts. During conversations about suicidal thoughts, self-harm, and eating disorders lasting up to an hour, none of these accounts, according to Common Sense Media, triggered a corresponding notification.

The organization also stated that ChatGPT missed more than a quarter of cases where its experts deemed it necessary to recommend a crisis service, medical professional, or other professional help to the teen. At the same time, Common Sense Media noted that some protective mechanisms worked properly - in particular, the chatbot mostly refused sexual roleplay content.

Researchers also saw problems in the educational mode. ChatGPT for Teens is supposed to encourage teens to solve tasks independently using hints and step-by-step explanations, but during testing, users could switch to the function of obtaining a ready answer. Common Sense Media also believes that Study Hours set by parents could be bypassed too easily.

OpenAI disagreed with the organization's assessment. The company stated that a significant portion of parental notification testing may have been conducted before the complete activation of the link between teen and parent accounts. This process can take several hours, so the company believes some results do not reflect the system's operation under normal conditions.

Common Sense Media responded that they clarified the readiness of ChatGPT for Teens features with OpenAI before starting the check. According to the organization, notification problems persisted even on some accounts that had been linked to parent accounts for much longer than the company's stated activation period.

OpenAI also explained that the ability to exit Study Mode during Study Hours is a deliberate product feature, not a technical error. The company intended this feature as a way to encourage the educational mode, not as a strict blocking of other ways to use ChatGPT.

ChatGPT for Teens was launched on August 18 as a separate set of settings and protective mechanisms for accounts of users aged 13-17. It is automatically activated for accounts the system identifies as teen, and includes enhanced content filters, learning tools, break reminders, and additional parental controls. OpenAI emphasizes that safety threat notifications are not a real-time monitoring system and may not detect every dangerous situation.

Common Sense Media recommends OpenAI to address the identified shortcomings and conduct repeated independent testing. OpenAI itself states that it will continue to improve the protection of minors and cooperate with experts.

Based on materials from: Common Sense Media, OpenAI, Bloomberg