How Thnkrr Checks What People Write
Thnkrr is a network for high school students. Everything someone writes on Thnkrr is checked before anyone else can see it.
What gets checked
- Posts
- Comments
- Messages
- The message a student sends when asking someone to be their mentor
- The "about" section and career goals on a profile
- A mentor's description of their experience
- A school's page description
Short details such as names, a college name or a job title can't include phone numbers, emails, links or social media handles.
What we look for
- Bullying or putting someone down
- Sexual content
- Signs an adult is trying to build an inappropriate relationship with a student
- Phone numbers, addresses, emails or social media handles, or trying to move a conversation off Thnkrr
- Hate
- Threats or violence
- Promoting drugs, alcohol, vaping or weapons
- Spam and scams
- Pretending to be someone else
- Signs that someone may hurt themselves
Moving a conversation somewhere we can't see is often the first step toward harm to a young person, so it isn't allowed anywhere on Thnkrr.
What happens after someone writes something
| What happens | What people see |
|---|---|
| It's fine — almost everything | It appears within moments. There's no sign it was checked. |
| A person needs to look at it | The writer sees that it's being checked. A post shows only to its writer until it's cleared. Comments and messages stay hidden from everyone, including the writer, so nobody can use them to test what gets through. |
| It isn't allowed | It never appears. The writer is told, in general terms, why it wasn't allowed and keeps their words so they can change them and try again. We never point to the exact words, so the rules can't be worked around. |
| It was already visible and is taken down later | A short note replaces it, so nobody thinks it disappeared by mistake. |
When someone edits their profile, the old version keeps showing until the new one is cleared.
Who does the checking
- AI reads everything first.
- Anything the AI isn't sure about, and anything suggesting someone may hurt themselves, goes to a person on the Thnkrr team.
- If the checking system is ever unavailable, nothing is published until a person has looked. Nothing goes out unchecked.
Extra care when an adult is involved
- The only adults students can talk with on Thnkrr are mentors, who are verified college students.
- Only a student can start a conversation with a mentor. Mentors can't contact students first.
- A student's first message to a mentor is checked before the mentor sees it.
- Messages between mentors and students are checked more strictly than messages between students.
- Every mentor conversation shows a notice that messages are checked. It can't be hidden.
- Nobody can delete messages. That way a student can't be pressured into deleting something, and nobody can erase what they sent.
If something harmful gets through
- Every post, comment, profile and conversation has a report option.
- Reports go straight to a person, not to the AI.
- The person who reported gets a confirmation. They aren't told what happened next, so a report doesn't become a source of conflict at school.
- Blocking is offered in the same place. Choosing "I don't know this person" also blocks them quietly.
If a student may be at risk
If someone writes about hurting themselves, the AI never simply blocks it. A person on our team looks at it. When we hold it back, we show the student the 988 Suicide & Crisis Lifeline.
When the checking gets it wrong
AI can make mistakes. Anyone whose writing is held back can change it and try again. For profile changes, they can also ask a person to take another look.
Students' privacy
- The AI sees only the words and a rough age group (13–15, 16–17 or 18+). It never sees a name, a school or an account.
- Any company whose AI checks students' writing must agree not to use it to train its AI, and to keep it for as short a time as possible.
What AI doesn't do on Thnkrr
- It doesn't guess how students are feeling or report that to adults.
- It doesn't decide what shows up in the feed. The feed is simply newest first.
- It doesn't pick who students should connect with or which mentor they should choose.
- It doesn't write anything for anyone.
- It never analyses faces.
What the checking is built from
There's no single standard tool for this. The biggest platforms for young people, such as Roblox and Discord, build their own. Platforms without large safety teams usually combine a few proven services. Here is each part, what Thnkrr uses today, and what other platforms use.
| Part | What it does | Thnkrr today | What other platforms use |
|---|---|---|---|
| Our own rules | Spot phone numbers, emails, links and social media handles instantly. No AI, and nothing leaves our servers. | Built, written by us | Every platform has these. Common free building blocks: libphonenumber-js, linkify-it, obscenity |
| A moderation service | A ready-made AI trained to spot common harms such as harassment, hate, sexual content and self-harm. | Not used yet | OpenAI's Moderation API (free; used by Khan Academy), Hive (used by Reddit and Yubo) |
| A child-safety specialist | AI trained specifically on grooming, sextortion and child exploitation. | Not used yet | Thorn's Safer (used by X and more than 80 platforms), or Hive's grooming model built with Thorn |
| A general AI model | Reads what's written against Thnkrr's rules, including context and slang. | Connected to Anthropic's Claude for now. The final choice isn't made. | OpenAI's models (usually through moderation companies), or open models such as OpenAI's gpt-oss-safeguard running on services that keep no data. Google's Gemini terms don't allow services used by under-18s. |
| People | Review anything unclear or serious, and every report. | Our own review page | Most platforms build their own review tools |
How we'll choose
- Test before trusting. We'll run the same set of realistic examples through each option and compare what it catches, what it wrongly blocks, how fast it is and what it costs.
- Protect students' writing. Any company we use must agree not to train on it, and should keep it for as short a time as possible.
- Prefer what's proven. Tools that other platforms for young people already rely on come first.
- Keep the list short. Every company we add receives students' writing and is named in school agreements.
For technical reviewers
- OpenAI Moderation API
- Hive Text Moderation
- Thorn Safer Predict
- gpt-oss-safeguard
- Roblox PII Classifier (contact-detail detection)
- Anthropic content moderation guide
- Full comparison — prices, licences, where each runs, data terms and who uses it: docs/15 §2.4
Not in place yet
- Choosing the AI services, by testing the options above.
- Consequences for repeat rule-breaking. Today, anything that breaks a rule is held back every time, but nothing yet happens to the account itself.
- Looking at a whole conversation for warning signs, not just one message at a time.
- A process for reporting serious child-safety cases to the authorities, as the law requires. This must be in place before the pilot opens to students.
Keep this page true to what Thnkrr actually does. When something in "Not in place yet" is built, move it up. The full detail lives in docs/12 and docs/15.