OpenAI officially launched the ChatGPT for Kids version for young users this week, with certain features restricted through an age recognition mechanism. However, the launch did not receive widespread approval, but instead sparked concerns from multiple child safety experts regarding the transparency and effectiveness of its safety mechanisms.
Experts in the fields of child safety and evaluation pointed out that the system must undergo extensive independent testing before being recommended to parents and teenagers. OpenAI also needs to improve transparency by clearly explaining how its safety mechanisms work. Robbie Tynjala from Common Sense Media stated that although OpenAI has made many important commitments, it must provide concrete evidence to prove the effectiveness of these features.
Josh Golin, Executive Director of the children's online safety initiative Fairplay, expressed more severe concerns. He pointed out that while the measures on paper seem to be moving in the right direction, there is a lack of corresponding accountability mechanisms. Looking back, OpenAI's restrictions on content related to suicide, violence, and drug use have previously failed or proved ineffective. Relying solely on the company's self-promises is not enough.
Experts strongly questioned the stability of the automatic age recognition system. Tynjala noted that OpenAI has not publicly shared data on the system's effectiveness, and neither the false positive rate nor the false negative rate is known. An automated system is unlikely to accurately identify all young users. In addition, teenagers are likely to try various methods to bypass age verification.
Regarding OpenAI's previous promise that "all flagged content will be reviewed by full-time employees and parents will be notified within one hour," experts remain skeptical. Tynjala questioned whether the automated classifier responsible for filtering content before sending it to humans is truly trustworthy; Golin called for more hard data to be disclosed, such as the number of conversations flagged daily, the size of the review team, and the average review time. It is difficult to believe that the company can hire enough staff to carefully review every dangerous conversation.
Experts agree that the mechanism requiring parents and children to actively link their ChatGPT accounts weakens the actual protective effect. Golin suggested that a more reasonable approach would be to automatically enable the protection mechanisms and integrate them into the product design, such as setting default usage time limits, rather than placing the entire responsibility of complex settings entirely on parents.
