Hour-long suicide chats on ChatGPT for Teens rarely triggered parent alerts, Common Sense Media found

A parent links their account to a teenager's and opts in to alerts about high-risk use. ChatGPT for Teens offers linked parents that option ^[2]. It is fair for the parent to assume the alerts start at that point. According to Tom Siegel, executive director of Common Sense Media's Youth AI Safety Institute, OpenAI disclosed after the tests that its systems can take several hours to activate on newly linked accounts ^[13].

Here's what teams tell themselves users do: link the account once, then rely on the alert. Here's what the test accounts, set up to behave like teens, actually did. On more than a dozen new accounts registered to 13- to 17-year-olds, testers spent up to an hour on suicide, self-harm or eating disorders, and only a handful of alerts reached the parent accounts ^[17]. Some safeguards were easy to turn off or get around ^[9].

OpenAI's defense rests on that lag. Spokesperson Eric Porterfield said "the bulk of their testing may have begun and concluded before activation of parental controls was complete, making their findings inaccurate" ^[4]. He added: "If that was the case, the tests would not establish whether parental safety notifications work as designed." ^[5] The objection is narrow. It argues the test could not settle the question either way. Siegel replied: "Some of our test accounts were linked within that window; however, others were linked for significantly longer, and still provided no notifications." ^[14] OpenAI's statement does not include its own figures on how often alerts fire ^[5]. Siegel has asked the company to share data with outside researchers ^[15].

In August the pitch was "stronger built-in safety protections" ^[2]. The report found the results uneven. Refusing sexual role play worked ^[8]. Clinical reviewers called much of the crisis content "clinically sound," but many conversations that should have led to a referral to professional help did not ^[3]. The weak points are the triggers: when to refer, and when to alert ^[17][3].

The report tested those triggers against California's Adam's Law and says ChatGPT failed tests tied to its crisis-referral, parental-alert and age-assurance requirements ^[10]. Gov. Gavin Newsom signed the law last month, and it takes effect in July 2027 ^[11]. Common Sense Media sponsored the bill and OpenAI endorsed it ^[12]. The institute is partly funded by the OpenAI Foundation. It says OpenAI reviewed a draft for factual accuracy but had no say over the findings ^[16].

In my view, a team shipping a chatbot to minors should keep parent alerts and crisis referrals out of its marketing until someone outside the team has tried to trigger them in long conversations. The cost is launch time, plus the risk of a public finding the team will want to contest, as OpenAI is contesting this one ^[6]. Siegel set a stricter bar for OpenAI: "A teen can spend an hour talking about self-harm without their parent getting a single alert. Until OpenAI fixes that and proves it with independent testing, ChatGPT should be for adults only." ^[18]

A team can sort its own safeguards on a two-by-two. One axis is whether a safeguard works in the first hours after a parent switches it on. The other is whether it still fires an hour into a single conversation. Works on both: it can go on the pricing page. Holds up in long chats but lags after setup: print the lag, in hours, on the setup screen. Works from setup but fails in long chats: keep it off the marketing page, because long chats are where the report found alerts rarely fired ^[17]. Fails both: call it a roadmap item.