Most parental controls for teens were built around one idea: let the parent see everything. Every text, every search, every message. That design breaks down with AI chatbots.
Teens talk to ChatGPT, Gemini, Claude and Character.AI about things they won't say out loud to anyone. Stress, friendships, their bodies, their sexuality, whether they're okay. If a parent can read those chats word for word, two things happen fast. The teen finds out, and the teen moves to an app the parent doesn't know about.
When we built TeenCue, we started from a different question: what is the least a parent needs to know to make a good decision? Our answer was the topic and how concerning it is. Not the words.
This post walks through how that works.
The output: topic and concern level, nothing else
For every message a teen sends on a supported AI chat service, TeenCue produces a small record:
- Service: which AI it was (ChatGPT, Gemini, Claude, Character.AI, Copilot and so on)
- Category and topic: what the message was about, at a level like "school", "emotional support" or "relationships"
- Concern level: how likely it is that a parent should check in
That record is what the parent sees in the dashboard, the iPhone app and the weekly Teen Brief email. Parents never see what was typed.
A typical brief reads something like: TeenCue noticed 1 signal that may be worth a conversation. It was related to leaning on AI for emotional support. That is enough for a parent to know it's time to talk, and it leaves the teen's words with the teen.
Rating the message
To produce that record, the message has to be read by something. We use an AI rating model from Anthropic, the company behind Claude.
The flow is short:
- The teen sends a message to an AI chat service in their browser or in Safari on their iPhone.
- TeenCue sends that message securely to the rating model with a prompt that asks for one thing: a structured rating (category, topic, concern level), not a summary and not a quote.
- TeenCue stores the rating and discards the message text.
Asking for a structured rating instead of free text matters. A summary is just the chat in shorter form. A topic plus a level is something a parent can act on without learning what their kid actually said.
What we keep and what we don't
We keep:
- The rating (service, category, topic, concern level)
- Counts and timing, so a parent can see things like "24 AI chats across 3 services this week"
We don't keep:
- The text the teen typed
- The AI's replies
- Other websites or browsing history
We're also upfront about the one place text does exist briefly. Anthropic, the rating provider, deletes the message within 30 days, except for chats its own systems flag for safety, which it can keep for up to 2 years. That is Anthropic's retention policy, and we'd rather say it plainly than hide it behind "never stored."
Ratings can be wrong
A model rating a single message without context will sometimes get it wrong. "I want to die" means something very different after a math test than it does at 2am.
So TeenCue treats a rating as a reason to talk, not a verdict. The product language is deliberately soft ("may be worth a look") and the parent view shows patterns across a week, not a single alarming line. The goal is a calm conversation, not a confrontation built on one flagged message.
Designed to be used openly
TeenCue does nothing until it's paired with a parent account using a short code, and we tell parents to let their teen know it's installed. That's not just ethics. Monitoring that relies on secrecy fails the moment it's discovered, and teens always discover it. Monitoring a teen knows about, and that doesn't expose their words, is something they can live with.
Where it runs
- iPhone: Apple's Screen Time APIs plus a Safari extension, no VPN or MDM. We wrote about that in How we keep AI chatbots inside Safari on a teen's iPhone.
- Computers: an extension for Chrome, Edge, Firefox and Brave.
If you're a parent, you can try it free for a week at teencue.com. If you're an engineer working on privacy-preserving safety tools, we'd like to hear how you'd approach this differently.

