OpenAI launched a teen version of ChatGPT this week, designed specifically for young users and using age-detection mechanisms to restrict certain features.
In response, Engadget interviewed several child safety experts. Their assessments of ChatGPT's teen version were mixed, but they all agreed that the system needs extensive independent testing before it is recommended to parents and young people. OpenAI must be more transparent about how its safety mechanisms actually work.
Robbie Torney, head of AI and digital evaluations at Common Sense Media, which provides families with recommendations on entertainment and technology products, said: “OpenAI has made some very important commitments this time, but it must provide evidence that these safety commitments and features actually work.”
Josh Golin, executive director of the child online safety organization Fairplay, took a harsher view: “On paper, some of the measures OpenAI announced do appear to be moving in the right direction, but there are still many reasons for concern when you look more closely, including the lack of accountability mechanisms. Social media has already shown that simply trusting companies is not enough to protect children, and OpenAI has faced similar problems. Before this, OpenAI already had content restrictions targeting suicide, violence, and drug use, but those restrictions either did not work or gradually degraded over time.”
Both experts questioned whether OpenAI’s automated age-detection system could reliably identify teenagers and switch them to the teen version of ChatGPT. Torney said: “OpenAI has not released performance data for this system. Its false-positive and false-negative rates are unknown, and an automated system is unlikely to identify every teenager on ChatGPT. When OpenAI previously planned to introduce certain adult-content features, one reason reportedly cited for their eventual cancellation was that age-estimation technology could not accurately identify enough teenagers.”
Torney also believes that teenagers will almost certainly try to bypass age detection and look for increasingly unexpected ways to evade it.
OpenAI’s most specific commitment is also the one experts are most skeptical can be implemented. Lauren Jonas, OpenAI’s head of youth and family affairs, previously said: “All flagged content will first be reviewed by full-time OpenAI employees before parents are notified. Our goal is to notify parents within one hour after a user submits a prompt.”
She also made clear that OpenAI has enough staff and processing capacity to ensure that all flagged content receives human review and that parents are notified within one hour.
Torney questioned this: “This process still depends on automated decisions, so the question is: Can the automated classifier responsible for screening content and then sending it to human reviewers really be trusted?”
Golin added: “I would like to see hard data, such as how many conversations are flagged and reviewed each day, how many people are responsible for reviewing them, and how long each review takes on average. It is difficult to imagine that OpenAI has hired enough people to carefully examine every potentially dangerous conversation.”
The three experts also unanimously agreed that requiring parents and children to actively link their ChatGPT accounts weakens the feature’s practical effectiveness. Golin said: “The vast majority of parents simply will not use parental supervision tools, partly because these tools are often designed to be difficult to find and difficult to operate.”
He believes a more effective approach would be to enable protections by default and build them into the product design from the outset: “Start with the strictest protection settings, such as default time limits, and then let parents decide whether to change them, rather than placing the responsibility on parents and requiring them to configure the limits themselves.”