ChatGPT Rated Unacceptable Risk for Teens After Parental Alerts Failed

What happened

An independent audit has rated ChatGPT an "unacceptable risk" for minors after OpenAI's teen safety features failed to perform as intended. The assessment was carried out by the Common Sense Media Youth AI Safety Institute, which ran more than 4,000 test prompts against the service, as reported by The Decoder.

The most serious finding concerns crisis conversations. On test accounts, discussions of suicide and self-harm did not trigger a parental alert, meaning the safeguard designed to bring a guardian into a high-risk moment did not activate in the scenarios tested. The institute has called for teenagers to be kept off the service until its safety can be independently verified.

The findings land at a time when AI chatbots are increasingly used for advice, companionship and schoolwork, and when the protections built around them are coming under closer scrutiny. The institute's position is a direct challenge to the assumption that existing teen settings are sufficient on their own.

Why it matters

For organisations deploying conversational AI, this story is a reminder that a safety feature is only as strong as its behaviour under real conditions. A control that exists on a product specification sheet, but fails to fire when a vulnerable user needs it, is a service failure, not a configuration detail. Testing has to cover the moments that matter most, not just the average interaction.

It also shows that independent evaluation is becoming part of the trust equation. Leaders building AI into customer-facing or public-facing services should expect that their safeguards will be tested by parties outside their own organisation, and should design for that scrutiny from the outset rather than treating it as a reputational surprise.

By the numbers

• More than 4,000 test prompts were used in the Common Sense Media Youth AI Safety Institute's assessment of ChatGPT.
• Zero parental alerts were triggered during the test conversations about suicide and self-harm on test accounts.

The Renascence take

Most coverage will frame this as a verdict on one company's product. The more useful reading is that the failure sits in the gap between a feature being built and a feature being proven to work in the moment of need. Parental alerts only protect anyone if they fire at the right time, and the audit suggests that gap is wide enough to matter.

Teams tend to measure safety features by whether they exist, not whether they respond under pressure. The principle that should guide every operator is that a safeguard is a service promise, and it must be tested against the highest-stakes journeys first, not the easiest ones. If you cannot show that your escalation path works for a distressed user at 2 a.m., you do not yet have a safeguard. Map the worst moments in your customer journey, test them end to end with independent eyes, and publish what you find.