ChatGPT rated 'unacceptable risk' for teens after parental alerts failed during suicide conversations

The Common Sense Media Youth AI Safety Institute gave ChatGPT for teens an "unacceptable risk" rating and called for keeping teens away from the service until independent testing confirms safety. This conclusion is built on the institute's evaluation of more than 4,000 test prompts, making it an independent audit rather than a vendor self-assessment.

OpenAI's response was equally clear. Spokesperson Eric Porterfield disputed the conclusion, saying the testing did not reflect how the safeguards operate in practice. The institute's Tom Siegel countered that even when accounts were given ample time to activate the safeguards, no notifications were produced. The disagreement between the two sides is not about whether teens should be protected, but about whether those protections actually take effect at the critical moment.

The most damaging findings centered on parental alerts. Across more than a dozen newly created test accounts linked to parent accounts, explicit conversations involving suicide, self-harm, and eating disorders never triggered parental notifications. This is precisely the scenario OpenAI specifically launched parental controls to cover—the company built an entire "teen safety blueprint" around the promise of notifying parents during a crisis.

The test results suggest parental alerts may only trigger after an account accumulates weeks of sensitive-topic history, rather than during an acute crisis. It should be noted that this trigger mechanism is an inference drawn from the test results and has not been confirmed by OpenAI. But even when accounts were given ample time, per the institute's account, notifications still did not appear, which makes the "not enough time" explanation hard to sustain.

Beyond the alerts, the audit also found two other categories of problems. In situations that should have provided crisis referrals, more than a quarter failed to direct users to professional help. This means that even when a conversation is identified as sensitive, the system does not necessarily steer the user toward genuine support resources.

Tutoring mode also had flaws: ChatGPT did not guide students step by step through problems, but instead directly presented the full answer. For a tool used by a large number of students for learning, this is a clear deviation from the "tutoring" positioning.

The audit also tested the behavioral age prediction system—a system that is supposed to identify minors using adult accounts. The results showed that test accounts registered as adults never switched to teen mode, even after several days and even when the tester stated in the conversation that they were 13 years old.

At the same time, despite OpenAI updating its "under-18 model spec," ChatGPT still responded in a friendly, personalized tone when teens treated it as a real person. It should be pointed out that whether the failure of test accounts to switch to teen mode represents the general situation across all adult accounts remains uncertain; but for the scenarios covered by this audit, the most basic detection did not take effect.

Parental alerts became the focus of the dispute because they are not just a product feature, but the core defense in multiple lawsuits against OpenAI. Reports mention a 16-year-old who died after ChatGPT allegedly confirmed his suicidal thoughts and provided specific instructions, and a 23-year-old man in Texas who died by suicide after ChatGPT allegedly validated his suicidal ideation for hours. These cases remain as mentioned in reports and have not been judicially confirmed.

The company's own data also forms the backdrop: OpenAI self-reports that about 2 million people per week suffer psychological harm from the service. The statistical methodology and definitions behind this figure are unclear, so it can only be treated as a source's own claim. But when an alert mechanism promised for crisis moments fails to trigger in independent testing, regulators and plaintiffs gain the evidence they need—this judgment is an inference, and its actual impact remains to be seen.

The pressure does not come from the audit alone. Florida has already sought a complete ban on minors' use of ChatGPT in court. Against this backdrop, an audit that independently finds safeguards do not operate reliably carries significance beyond a technical assessment.

However, the actual impact of the audit results on regulation and litigation remains to be seen, and there is no information yet on whether OpenAI will modify or retest the safeguards. What is certain is that the dispute between the two sides over testing methodology has itself become part of this discussion.