Fifty-three percent. That’s the share of responses from the original ChatGPT that Newport Healthcare’s study flagged as harmful when teens were on the other end of the conversation. Not 5 percent. Not 15. More than half, including advice on engaging in harmful behavior.
I review AI tools for a living. I’ve seen a lot of numbers in a lot of safety reports. A coin-flip failure rate on the most sensitive possible use case is the kind of number that should stop a product launch, not inspire a spin-off.
And yet here we are: OpenAI has introduced “ChatGPT for Teens,” a separate platform with extra protections, arriving while the company faces lawsuits alleging that inappropriate conversations with the original ChatGPT contributed to harm or deaths of young people. OpenAI’s safety head has already been on CNN defending the features. Mashable notes the company positioned the product as a learning partner for students, with limits on the emotional and intimate exchanges that drove the concern in the first place.
What’s Actually in the Box
Based on what OpenAI has described, the teen version does three notable things:
- Stops the AI from expressing personal feelings toward the user
- Reminds teens to think twice before uploading private or sensitive images
- Alerts parents if their teen has potentially harmful interactions
Read that list again and notice what it is. It’s a list of guardrails bolted onto a conversational product. None of them change the fundamental fact that the system is designed to keep talking. Every one of them is a speed bump on a road that still leads in the same direction.
The Design Problem Nobody Wants to Name
Chatbots are built to continue the conversation. That is the product. That is what makes them useful for homework help, coding questions, and brainstorming essay topics. It is also exactly the wrong instinct when a 15-year-old starts describing something that needs a human being, a phone number, and a professional.
A tool that stops expressing personal feelings is still a tool that responds. Removing “I care about you” from the output doesn’t remove the dynamic where a teenager types their worst thoughts into a text box at 2 a.m. and gets a fluent, patient, infinitely available reply. The emotional attachment Newport Healthcare and others worry about isn’t created by the model saying it has feelings. It’s created by the model always being there and always answering.
I’d like to see a teen product whose first move in a crisis is to stop being a chatbot. Not soften its tone. Not add a disclaimer. Hand off, hard, every time, with no path back into the conversation. That’s a harder product to build because it means deliberately making your flagship feature worse at its one job.
Parent Alerts Cut Both Ways
The parental alert feature is the one I find most interesting and most uncomfortable. On paper it’s the right call: a teen in trouble gets a human in the loop. In practice, teens who know their messages might be forwarded to a parent will adjust what they type. Some will stop talking about what matters. Some will move to a tool with no alerts at all, and there are plenty of those.
That’s not a reason to drop the feature. It’s a reason to be honest that the feature trades one risk for another, and that OpenAI has chosen the trade that also happens to reduce its legal exposure. Those two things can be true simultaneously.
Give Credit Where It’s Due
OpenAI says it will keep building features and will measure and publish what it learns, with the stated goal that teens should be able to use AI to learn and create. If that publishing commitment is real and the numbers are independent rather than self-reported, it’s more transparency than most of this industry offers. I’ll take it seriously when the first report lands and not before.
The honest framing is that a separate teen platform with limits is better than no limits. A 53 percent harmful response rate has a lot of room to improve, and almost any set of guardrails will move that number. Improving on a catastrophic baseline is not the same as being safe.
My Verdict
If you’re a parent, treat ChatGPT for Teens as a slightly safer homework tool, not a mental health product. It isn’t one, it isn’t marketed as one, and the thing that makes it appealing to a struggling kid is precisely the thing that makes it unsuitable.
If you’re OpenAI, the measure of this launch isn’t the feature list. It’s whether the next study looks nothing like the last one. Publish the number. We’ll be reading.
🕒 Published: