Independent testing finds gaps in ChatGPT’s teen safeguards
Common Sense Media’s assessment found gaps in ChatGPT’s protections for teenage users after roughly 4,000 interactions. Parental alerts arrived late and study mode was easy to bypass. Education Week obtained a response from OpenAI, which disputed the testing method and cited technical errors affecting safeguards. The findings concern the tested experience and sharpen questions about protections offered to students and parents.
Artificial Intelligence··Night
An independent assessment finds gaps after thousands of interactions
Common Sense Media, a nonprofit that studies young people and technology, assessed ChatGPT’s teen protections through roughly 4,000 interactions. It rated the service an unacceptable risk for users under 18. Education Week reported the findings and interviewed assessment lead Robbie Torney. The assessment tested the experience offered to teenagers, including parental warnings and a study mode intended to support schoolwork.[1], [2]
Parental warnings were late and study mode was bypassed
Testers posing as teenagers spent up to an hour discussing dangerous subjects on accounts linked to a parent, without the parent receiving a warning. Common Sense found that older accounts with weeks of sensitive conversation history could prompt alerts.[1]
Study mode also offered a way to show an assignment’s answer, and teenagers could switch the feature off. These findings concern the experience examined in the assessment.[1]
OpenAI disputes the method and cites technical problems
OpenAI responded to Education Week and disputed the testing methodology. The company said technical errors prevented some safeguards, including parental notifications, from activating properly. It said the testing did not accurately reflect the teen safeguards in practice. OpenAI also said study mode was developed with feedback from child-development and learning-science experts and teenagers. Common Sense’s assessment lead Robbie Torney objected that an answer button lets pupils bypass the intended learning process.[1]
