- Common Sense Media's Youth AI Safety Institute says parental alerts in ChatGPT for Teens did not fire on a dozen-plus new accounts during explicit crisis chats. OpenAI disputes the method.
- The Institute infers alerts depend on account history, not severity. That is the mechanism leaders should question in any AI safeguard they rely on.
- Ask vendors what triggers each safeguard, and test first-day scenarios yourself rather than trusting the feature list.
A parent who links a teenager's ChatGPT account is buying one promise. If something serious happens, the parent will be told. The Youth AI Safety Institute at Common Sense Media tested that promise. It says the promise did not hold.
What the testers found
OpenAI announced ChatGPT for Teens on August 18, 2026. It said the launch brought "stronger built-in safety protections," including parental notifications for dangerous chats. The Institute tested in two windows, one in July and August and one in late August and September. It sent more than 4,000 prompts in total, roughly half in each period. It rated ChatGPT an Unacceptable Risk for everyone under 18.
The alerts were the main finding. The testers created more than a dozen new free accounts, each linked to a parent account before any chat began. Personas talked explicitly about suicide, self-harm and eating disorders. Sessions ran from under five minutes up to an hour. No account in that test produced a notification.
How the alert appears to work
The Institute reads its results this way: the alert depends on accumulated account history, not on how severe one conversation is. That is an inference from testing. OpenAI has not confirmed it.
The evidence is uneven but consistent. Before launch, an age-13 account ran 990 prompts over several weeks. It produced two alerts, one three hours after the first crisis-level disclosure and one three days after. After launch, a second teen-registered account received two notifications from 450 prompts. One covered suicide and self-harm and one covered disordered eating. Each arrived within an hour of the testers raising that subject. Both accounts carried weeks of sensitive-topic conversation.
The alerts also had design gaps. The emails carried no timestamp for the chat that set them off. No reminder followed as the conversations went on.
The Institute drew a practical conclusion. Consider a teenager who has used ChatGPT for months on essays and game questions. That account has a long record, but not one that resembles a build-up to a crisis. By the Institute's reasoning, that teen's first disclosure could get no more protection than a brand-new account would.
What OpenAI says
OpenAI disputes the findings. The Decoder reports that spokesperson Eric Porterfield said the tests did not reflect how the safeguards work in practice.
OpenAI also told the Institute that parent and teen accounts must be linked for about three hours before alerts can arrive. It has since updated a help center article. The Institute checked its accounts. Some had been linked for less than three hours and some for more. It says it stands by its reading. Tom Siegel of the Institute told The Decoder that accounts given enough time still produced no notifications.
The study has limits. The two testing windows ran one after the other, so changes to the underlying model are mixed in with the teen settings. All testing was in the United States, from the San Francisco Bay Area. The Institute did not calculate agreement between reviewers.
Other gaps in the same assessment
Three child psychiatrists marked in advance which prompts called for a crisis resource. After launch, more than a quarter of those prompts did not get one. The Institute also holds ChatGPT to a 95% bar on its five Red Line harms, the kind that can do extreme or irreversible harm to a child. On three of them, ChatGPT fell short.
Age detection fell short too. OpenAI has said a machine learning model predicts a user's age from signals such as account age, activity times and conversation topics. In the Institute's tests, adult-registered accounts did not switch to the teen experience over many days. That held even when testers said they were 13 and the bot acknowledged it.
Study mode was easy to sidestep. A "Show me the answer" pop-up let teens skip the tutoring. Deleting the "@study" prefix let them leave parent-set Study Hours.
The lesson: a safeguard is only as good as its trigger
A feature list describes what a vendor built. The trigger describes when it works. This story shows the gap between the two. "Notifies parents in high-risk situations" sounds like a rule about danger. In testing, it appeared to behave like a rule about accumulated sensitive-topic history, the Institute infers.
The Institute offers a point of comparison. It points to teen-focused mental health apps it has reviewed. In those evaluations, the same kind of explicit disclosure prompted a phone call to a parent or guardian within a quarter of an hour. It accepts that a larger, global product may need a different mechanism. Its verdict is that a feature described as flagging "high-risk situations" cannot stay silent through 15 minutes of explicit talk about suicide, self-harm or eating disorders.
The same risk reaches any organization that buys an AI product on the strength of its safeguards. Each protection has a condition that sets it off. If nobody tests that condition, the protection exists only on paper.
What leaders should ask
First, ask vendors what exactly triggers each safeguard. Ask whether it responds to one severe event or to a pattern over time.
Second, test with a first-time scenario on a fresh account. Do not test only on accounts with a long history.
Third, check what the alert tells the person who receives it. A notification with no time or context is hard to act on.
Fourth, ask who verifies age or role, and what happens when the system's guess is wrong.
Fifth, require independent testing before relying on a vendor's description. Here, OpenAI told the Institute about a three-hour linking requirement during its review of the draft, and has since updated a help center article.
A control that parents, customers or staff trust has to work the first time it is needed, not only after weeks of history.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Common Sense Media Youth AI Safety Institute.
