Why we stand behind our ChatGPT findings

By Robbie Torney

Head of AI & Digital Assessments, Youth AI Safety Institute

On Oct. 7, the Youth AI Safety Institute rated ChatGPT for Teens an Unacceptable Risk for kids. OpenAI told several media outlets that the "bulk" of our testing "may have begun and concluded before activation of parental controls was complete."

That's not what happened.

Before we began testing, we confirmed with OpenAI that its newly added Teen mode features, including Study Mode and parental alerts for conversations about eating disorders, were fully launched on our accounts. Almost all of our testing ran on accounts that had been in use for weeks, not hours. On those accounts, testers worked through hundreds of prompts that child and adolescent psychiatrists had flagged in advance as warranting crisis resources.

We also ran a separate, targeted test to answer one important question: What does it actually take to trigger an alert to a linked parent account? We set up brand-new accounts, linked each one to a parent account before the first chat message, and had testers disclose suicidal thoughts, self-harm, or disordered eating, in conversations lasting up to an hour. None of those accounts produced a notification.

After we finished our tests and asked OpenAI to comment, the company disclosed that alerts can take up to three hours to switch on after accounts are linked. So we checked our logs. Some of our alert tests ran on accounts linked for less than three hours. But others ran on accounts linked well over three hours—past the delay OpenAI described. No parental alerts arrived either way.

Nothing in OpenAI's disclosure changes our conclusion: Parental alerts are unreliable in a crisis.

The delay is itself a gap in ChatGPT for Teens' safety net. In our testing, accounts were labeled linked in the parental control interface immediately. OpenAI never informed parents that safety alerts could take hours to start working. Only after our tests did OpenAI update its Help Center with a warning.

When OpenAI tells us it has made improvements, we'll test ChatGPT for Teens again and publish what we find. We'll rerun the same tests from this assessment, including crisis notifications and other key safety functions.

Here's what we'll be looking for on parental alerts:

• Protections that are on from a teen's very first conversation.

• Clear notice to parents, during setup, about when alerts are live.

• Published criteria for what triggers an alert, and how quickly it should arrive.

• Crisis alerts that arrive reliably and quickly, including the first time a teen discloses a serious crisis.

If ChatGPT gets better, we'll say so. That's how independent testing is supposed to work: We test, companies fix what we find, we test again, and families end up with products that are safer and that they can trust.