ChatGPT for Teens has special safeguards. A watchdog group finds most don't work

According to NPR’s report on the watchdog group’s findings, ChatGPT for Teens was marketed with special safeguards—but most of them don’t actually work. That’s not just a technical failure; it’s a philosophical one. We’ve created an AI that can sound like the world’s most understanding friend, yet it lacks the instinct to sound the alarm when that friendship matters most. It’s as if we’ve handed a teenager a diary that writes back, but the diary has no sense of duty when the words turn dark. The deeper weirdness here is that AI companions are designed to mirror human warmth. They’re trained on billions of conversations, absorbing the rhythms of empathy, humor, and reassurance. But mirroring is not feeling. A chatbot that talks like a peer may earn trust precisely because it never judges, never panics, never calls a parent. For a teenager in emotional pain, that can feel like a lifeline—but if the safety net fails, it becomes a beautifully worded echo chamber with no exit. This also exposes a strange asymmetry in our relationship with machines. We expect human friends to recognize cries for help, even subtle ones. But we’re still figuring out how to program that recognition into a language model. Common Sense Media found that the safeguards simply don’t trigger in many cases involving self-harm and suicide. That means the AI’s “friendship” is an illusion of care, not a structure of support. In the wild landscape of adolescent psychology, that’s a dangerously convincing illusion. The bigger question is accountability. When a teenager confides in an AI, who is responsible for what happens next? The tech company? The parents? The algorithm itself? Right now, the answer is murky—and that’s the most unsettling part of this story. We’re running a massive, unregulated psychological experiment on a generation that has never known a world without conversational AI. The safeguards were supposed to be the guardrails. If they don’t work, we’re driving blind. Still, there’s something genuinely wondrous about the fact that we can build a machine that feels so human. But the wildest thing about AI isn’t that it’s smart—it’s that we hand it our hearts and expect it to know what to do. This report is a reminder that empathy, even simulated, must come with an ethical circuit breaker. Otherwise, the friend we built may be the one who listens to everything and tells no one.