OpenAI launches teen version of ChatGPT as experts question safety measures

OpenAI has launched a teen version of ChatGPT, designed specifically for younger users and featuring age verification to limit certain functionalities. However, multiple child safety experts have raised concerns about the transparency and effectiveness of the safety mechanisms. According to Engadget, these experts are divided in their opinions, but all agree that the system requires extensive independent testing before being recommended to parents and young users. Robby Torni, head of AI and digital assessment at Common Sense Media, emphasized that OpenAI must provide evidence that its safety promises and features are actually effective. Josh Goling, executive director of the Fairplay Children's Cybersecurity Initiative, expressed stricter concerns, noting that while some measures appear to be moving in the right direction, there are still significant issues, including a lack of accountability. He pointed out that previous content restrictions, such as those for suicide, violence, and drug-related content, either failed to work or became ineffective over time. Both experts questioned whether OpenAI's automatic age recognition system could reliably identify teenagers and switch them to the teen version of ChatGPT. Torni highlighted that OpenAI has not disclosed data on the system's accuracy, including false positives and false negatives, and that past attempts to introduce adult content features were abandoned due to the system's inability to accurately identify enough teenagers. OpenAI's most specific commitment, which also drew the most skepticism from experts, was the promise that all flagged content would be reviewed by full-time employees before notifying parents, with the goal of informing parents within an hour of a user's query. Torni questioned the reliability of the automated system responsible for filtering content and passing it to human reviewers. Goling also called for hard data, such as the number of daily conversations flagged and reviewed, the number of reviewers, and the average time spent on each review. All three experts agreed that the requirement for parents and children to actively bind ChatGPT accounts undermines the effectiveness of the feature. Goling noted that most parents do not use parental supervision tools, partly because they are often difficult to find and operate. He suggested that more effective measures would be to default to the strictest protection settings, such as time limits, and let parents decide whether to modify them, rather than placing the responsibility on parents to set these limits themselves.