It’s 11:40 on a Tuesday night. A fifteen-year-old is lying sideways across her bed, phone propped against a pillow, typing something she hasn’t said out loud to anyone. Not to her mom down the hall. Not to her best friend, who would probably screenshot it. She types it to a chat window, because the chat window is awake, it’s patient, and it never looks at her funny the next morning.
That moment is the whole reason OpenAI built ChatGPT for Teens, and it’s also the reason the product is so hard to get right.
What actually changed
OpenAI introduced ChatGPT for Teens as a separate experience with extra safety protections for younger users. It arrives against a serious backdrop: lawsuits alleging that inappropriate conversations with the platform contributed to mental health harm and to the deaths of young people. The company has positioned the new version as a learning partner for students, and said it limits the emotional and intimate exchanges between teens and AI that drove much of the concern in the first place.
Two specific changes stand out. The first is that the AI is designed not to express personal feelings toward a user. The second is a set of reminders nudging teens to think twice before uploading private or sensitive images. OpenAI has also said it plans to keep adding features and to measure and publish what it learns along the way. The company’s safety head has publicly defended the approach.
If you’re new to how these systems work, that first change is bigger than it sounds.
Why “I care about you” is a design decision
People often assume a chatbot’s warmth is a side effect of it being smart. It isn’t. It’s a setting. The tone, the follow-up questions, the little affirmations — those are shaped by instructions and training choices made by the company, and they can be turned up or down.
So when a model says something like “I’m really glad you told me that,” it’s not reporting an inner state. It’s producing the kind of sentence that keeps a conversation going. That’s fine when you’re stuck on a chemistry problem. It’s much more complicated when a teenager is in distress and the most engaging-feeling response is also the one that keeps them in the app instead of reaching a person.
Stripping out expressions of personal feeling is OpenAI’s attempt to break that loop. The AI can still be useful without pretending to be attached to you.
The part that still worries me
Research on AI chatbots and teen mental health is still limited, but what exists points in an uncomfortable direction: these tools can cause harm, reinforce dangerous ideas, and threaten well-being. Some cases have involved suicide. New guardrails do not erase that record, and a product update is not a study.
There’s also a structural tension worth naming. A chat product succeeds when the conversation continues. A mental health response often succeeds when the conversation moves — to a parent, a counselor, a crisis line, anyone with a body and a phone number. Those two goals pull in opposite directions, and no amount of careful wording fully resolves that. A system that is excellent at staying in the conversation is, by design, not excellent at ending it.
I can’t tell you from the available reporting exactly how the teen version behaves in a live crisis, moment by moment. That’s precisely the kind of detail that matters most and that outside researchers, not press releases, will have to establish.
What this means for parents and non-technical readers
You don’t need to understand transformer models to make good decisions here. A few practical framings help:
- Treat it as software, not a friend. The empathy is a feature, not a relationship. Saying that out loud to a teen is more useful than banning the app.
- Assume the default is engagement. Unless a company has specifically designed against it, chat products are built to keep you typing.
- Separate the use cases. Homework help, essay feedback, and study plans are reasonable uses. Late-night emotional processing is where the risk concentrates.
- Watch for substitution. The warning sign isn’t a teen using AI. It’s a teen using AI instead of people.
- Ask what gets published. OpenAI said it will share what it learns. Hold them to it, and look for independent research alongside it.
A reasonable step, not a resolution
ChatGPT for Teens looks like a genuine attempt to narrow a known risk, and the specific choices — no professed feelings, friction before sharing sensitive images, less intimacy by default — are the right category of fix. They address how the product behaves, not just what it’s allowed to say.
But the hardest problem isn’t a teenager asking an AI something inappropriate. It’s a teenager at 11:40 at night who finds it easier to talk to software than to a person, and a system whose instinct is to answer. Guardrails can make that exchange safer. Getting her to the person is still our job.
🕒 Published: