Keeping young people safe in AI conversations
How Cerebrellic responds when someone may be in crisis, why it never stonewalls them, and why sexual roleplay with anyone who appears under 18 is always blocked.
Young people talk to AI. They use it for homework, for advice they don't want to ask an adult for, and sometimes for company late at night. Most of those conversations are ordinary. Some are not. A teenager in crisis may tell a chatbot something they haven't told anyone else. And some products will happily follow a young user into sexual roleplay, because nothing told them not to.
Those two moments matter more than any others we check. This is how Cerebrellic handles them.
When someone is struggling
The easiest thing for a safety system to do with a message about suicide or self-harm is to block it. Most do. The person writes something painful and gets back a refusal.
We think that's the wrong answer. Someone who reaches out and gets "I can't help with that" learns that reaching out doesn't work. So Cerebrellic does three things instead.
It notices, without shutting the door. Every workspace has a Self-harm risk check that reads each conversation for signs that the person may be thinking about suicide or hurting themselves. It doesn't block the message. It flags it, and the model still answers, because a response is what the person needs.
It checks what the model says back. A separate check reads every answer before it's returned. It rejects answers that give methods or details that could be used for self-harm, encourage it, or brush off what the person said. When an answer fails, the model is told why and has to answer again. The person only ever sees an answer that has passed.
It never leaves someone with nothing. If no answer passes, we don't send a refusal. We send a short, caring reply. It acknowledges what they said, tells them they deserve support, and points them to people who can help right now. In the US, that's the 988 Suicide & Crisis Lifeline, which you can call or text on 988. Elsewhere, it's findahelpline.com for local services.
No sexual roleplay with minors. Ever.
Some AI products are built for companionship and roleplay. Many others can be pushed into it. Either way, when the person on the other side appears to be under 18, sexual or romantic roleplay has to stop. This isn't a setting.
Cerebrellic reads the whole conversation, not just the last message. If a user says they're 14, mentions their school year, or otherwise appears to be a minor, that stays in view for every message that follows. The Minor sexual content check blocks requests that steer towards sexual or romantic content with someone who appears to be under 18. It blocks answers that contain it.
It covers characters, too. Roleplay often runs through characters. If any character in a sexual scene is described or implied to be under 18, the content is blocked, whatever age the user claims to be.
It sits alongside our Child safety check. That check already stops messages that seek sexual or predatory contact with a real child, or try to get a child alone. Together they cover both the child in the conversation and the child being talked about.
Always on
These checks are on in every Cerebrellic workspace, including workspaces created before they existed. Child safety, Minor sexual content and the self-harm checks can't be switched off. Workspaces can make them stricter. They can't make them weaker.
Every response tells your app what happened. The lifeline.flags section names the check, its decision and its confidence, so your product can respond as well. For example, it can show support resources in your own interface, or lock a conversation for review.
What you should do
Cerebrellic protects each conversation. You know your users, so some steps are yours:
- Know if your users might be young. If your product could be used by under-18s, plan for it, even if your terms say otherwise. Age checks, parental controls and teen accounts all happen in your product, not ours.
- Act on self-harm flags with care. Have a plan for what your team does when a Self-harm risk flag appears. Never punish or shame a user for one. A suspended account is the last thing someone in crisis needs.
- Act on Minor sexual content flags quickly. Review them, restrict the account where it's warranted, and follow the steps in our article on protecting children in AI products if a child may be at risk.
- Don't build a product that depends on blurring the line. If your engagement depends on romantic roleplay, decide now how you'll keep minors out of it.
What we won't claim
- We can't verify age. We judge from what's in the conversation. Someone who never mentions their age, or lies about it, will look like an adult to us. That's why age assurance in your product matters.
- We are not a crisis service. Our checks make an AI's answers safer and point people to real help. They don't replace trained people.
- Checks can miss. No automated judgement is perfect. We test these checks against hard cases and keep improving them, but they're one layer of protection, not the whole system.
Why we built it this way
It would be easier to block everything that sounds difficult and call it safety. But a teenager who is struggling deserves to be heard, not refused. A teenager who is being drawn into something they shouldn't be deserves a hard stop, not a policy setting. We built Cerebrellic to tell those two moments apart and to get both of them right.
If you're struggling right now, you don't have to go through it alone. In the US, call or text 988. Anywhere else, visit findahelpline.com to find support near you.