ChatGPT's age estimator left testers who said they were 13 outside its teen protections

Common Sense registered its test accounts to 19-year-olds, above the 13-to-17 range that sends a declared user into the teen experience ^[3][4]. That left OpenAI's age estimator as the only route into teen mode the test could exercise ^[14]. The input was about as strong as a self-report gets. Over seven days, according to Common Sense, testers supplied adolescent context and said outright that they were 13 ^[4].

ChatGPT sometimes repeated that age back in its replies, and the teen protections still stayed off ^[5]. The stated age reached the text the model wrote without changing the account's state ^[5]. Two designs would produce that result. The estimator may not read conversation content at all, or it may read it and give a stated 13 too little weight to override a registered 19.

The parental-alert dispute is about timing. The assessment records OpenAI's explanation that linked accounts need about three hours before alerts begin working ^[18]. A crisis conversation of up to an hour, started soon after linking, would end inside that window ^[15]. The Institute says some tested accounts had passed the three-hour mark and stands by its conclusions ^[18].

OpenAI's own figures measure activity. The company said teens with access sent "about 2.7 million more learning-related messages on average" than teens still waiting for access ^[9]. It also said more than 80% of teens with sessions longer than three consecutive hours had at least one learning-related prompt ^[9]. One such prompt in a three-hour session is enough to count ^[16]. The announcement does not explain the averaging period or how the two groups were compared ^[10].

Common Sense is careful about scope. Its tests used US-based accounts, left out some features, and could not separate changes to the underlying model from changes to the teen settings ^[8]. The findings identify failures under tested conditions; they are not an overall failure rate for ChatGPT for Teens ^[8].

For the age result to transfer to another product, that product would need the same shape: protections keyed to an account-level age state, and nothing said in a conversation able to change it. On this test's evidence, OpenAI's teen experience has that shape ^[3][5]. I'd apply the finding to teams that gate protections the same way, and to nobody else until they are tested.