Child safety experts have warned that the security features in the teen-focused version of ChatGPT, recently released by OpenAI, may not be enough to protect them.
The release of the version that automatically directs users under 18 to an experience tailored specifically for them came a year after the system's development was announced, which coincided with a lawsuit filed by the parents of a 16-year-old teenager accusing ChatGPT of inciting his suicide.
The new version includes educational features such as "Study Mode" and graphical visualizations, with notifications for parents via linked accounts when unsafe conversations about eating disorders are detected, and an updated template guide that confirms that ChatGPT "should not use romantic language, encourage emotional dependency, or suggest that it has feelings or consciousness."
Robbie Turney, head of AI at Common Sense Media, said, "The announcement contains important commitments, but we need evidence that these commitments and features actually work." He added that the automated age-determination system doesn't detect 100% of teenagers, noting that OpenAI had previously postponed features related to adult content due to the inaccuracy of its age-estimation technology.
Josh Golin, executive director of the nonprofit Fairplay, was more critical, warning that the "lack of accountability mechanisms" is worrying and that "trusting companies doesn't protect children, as we've seen with social media and OpenAI itself." He pointed out that security features requiring parents and teens to link their accounts are often difficult to use, reducing their effectiveness.
In response, OpenAI pledged to have all reported content reviewed by full-time staff and to notify parents within an hour of the concerning conversation. However, experts questioned the company's ability to deliver on this promise, noting that previous tests showed delays of up to 48 hours in notifications.
Turney said: "The way teenagers express distress varies across cultures, requiring a global team capable of dealing with them in different countries."
Experts agreed that the requirement to link parents' and teenagers' accounts is a weakness, as parents rarely use parental control tools, either because they are difficult to find or because they are complicated to use, pointing to reports that showed more than half of the security features on platforms like Instagram, Snapchat, TikTok, and YouTube failed.
OpenAI said in a statement to Engadget: "Our goal is to support users and help them connect with real support resources, and we will continue to test and enhance this protection with experts, families, and teens." However, the company did not specify what actions it would take if it detected malicious conversations from unlinked accounts.
