safety doc · last reviewed october 2026 # Safety isn't a setting. It's the spine of how SafeMeet works. Every message is moderated before delivery. No exceptions, no toggles, no paywalls. the promise ## Four claims. All testable. Moderated before delivery Every message runs the pipeline before it reaches the recipient: inline policy rules first, then, for anything they don't block, the AI classifier and an image scan when there is an image. Sender doesn't opt in. Recipient doesn't opt out. Anonymous by default No email as a guest (a signed-in account keeps its sign-in email). No phone. No public profile. Guest sessions are first-class. We hash a soft device fingerprint server-side for sanction carry-over — raw IPs and device data are never stored. Fail-closed, always If the moderation provider is unreachable, messages stop sending. A message sent then is withheld from the recipient and is not queued to go out later. Honest about limits AI moderation is a floor, not a ceiling. Sarcasm slips through. Slow manipulation slips through. We publish what we catch, what we don't, and what to do when we miss something. behind every message ## What happens when you press send. Four steps. Two outcomes. No exceptions. [How AI moderation works](https://safemeet.now/guides/how-ai-moderation-works) Send user types & sends Moderate filter, then AI Decide safe? threshold met? Deliver recipient receives Block not delivered Safe path Most messages clear the pipeline fast enough that you won't notice the check. You won't feel the moderation, you'll just feel the conversation. Blocked path A message the check scores high enough to block is not delivered at all. Fail-closed: if moderation can't run, sending stops. fail-closed ## When the check can't run, nothing is delivered. If the AI check can't return a verdict, for example because its provider and the fallback both return an error, time out or give an answer we can't read, the message is withheld from the recipient and is not queued to go out later. — 01 AI provider returns 5xx or times out Retry once · then try the fallback provider · if it fails too, the message is withheld: not delivered, not queued — 02 Rate-limit on the provider Retry once · then try the fallback provider · if it is limited too, the message is withheld the same way — 03 Extended outage Every send fails closed (blocked, not queued) until moderation recovers honest limits ## AI moderation is a floor. Not a ceiling. No automated system catches everything. We don't pretend ours does. Here's what we catch, what we miss, and what to do when we fail. [Read the full safety doc ](https://safemeet.now/safety) What AI catches - A phone number or home address shared in a chat: in our evals the check flags it; a flagged text is delivered with a safety notice to its sender, and one the check scores high enough to block is withheld - Direct threats of violence: scored against the violence category - Slurs and named hate speech: scored against the hate category - Harassment aimed at the other person: in our evals the check flags it; a flagged text is delivered with a safety notice to its sender, and one the check scores high enough to block is withheld - Sexual content directed at strangers: scored against the sexual category AI still misses - Sarcasm and dark humor read as serious - Coded language and emerging slang - Slow-build manipulation across many messages - First-time patterns we haven't seen yet - Context-dependent harm ("send me a picture" of what?) - Things a human friend would clock instantly If the AI missed it, here's what to do. In a chat with a person, Panic ends the chat and files a report in one tap. Once a chat with a person ends, you can report it from your inbox. You can always leave. Sanctions carry by fingerprint, even if they clear cookies. Panic Report Block End dangers playbook ## Eight patterns, eight responses. We name the harm we expect to see, and we name how we respond. If a pattern isn't here, we publish it the moment we see it. — 01 Off-platform contact requests Anything asking for phone, email, or social handle. Scored against the personal-data category on the pre-delivery pass. In our evals a phone number or home address shared in a chat is flagged; a flagged text is delivered with a safety notice to its sender, and one the check scores high enough to block is withheld. A phone number, email address, link, @handle or crypto address it recognises in a delivered message is also hidden from the person who receives it. [How contact details are hidden](https://safemeet.now/guides/does-safemeet-hide-your-phone-number) — 02 Excessive compliment + escalation Classic grooming pattern — heavy flattery, fast escalation to private contact. Each message is checked on its own: a strong enough grooming signal blocks it, but the check does not read the messages before it, so a slow build across many can get past it. — 03 Minor-safety language signals School references, grade level, supervised time-of-day. A periodic scan scores a batch of recently active chats and logs only a medium or high score — never reviewed by a person. — 04 Crypto / investment pitches Scored against the scam category on the pre-delivery pass; a flagged text is delivered with a safety notice to its sender, and one the check scores high enough to block is withheld. A Bitcoin or Ethereum wallet address it recognises in a delivered message is hidden from the person who receives it. Cross-user duplicate detection is written but not yet wired into the pipeline. — 05 Targeted harassment Slurs, threats, named hate speech. Scored against the hate, violence and harassment categories on the pre-delivery pass. In our evals a targeted insult aimed at the chat partner is flagged; a flagged text is delivered with a safety notice to its sender, and one the check scores high enough to block is withheld. — 06 Image / photo pressure An image is scanned before delivery, and one that cannot be confirmed safe is held rather than sent. Text pressuring for photos is scored as sexual or exploitation pressure; a flagged text is delivered with a safety notice to its sender, and one the check scores high enough to block is withheld. — 07 Identity / number harvesting Phone numbers, social handles, payment IDs. Moderation scores them against its personal-data category before delivery. In our evals a shared phone number or home address is flagged; a flagged text is delivered with a safety notice to its sender, and one the check scores high enough to block is withheld. The phone numbers, @handles and emails it recognises in a delivered message are hidden from the person who receives it; it has no pattern for a payment ID or a street address. [Does SafeMeet hide your phone number?](https://safemeet.now/guides/does-safemeet-hide-your-phone-number) — 08 Public-figure impersonation Claims of being a celebrity or influencer go through the same AI moderation as everything else. No separate human queue — there isn't one. ## Someone says they want to hurt themselves Disclosing distress is never punished here, and the crisis resources you are shown are published, not improvised. The protocol covers chats with people and chats with the AI companion alike. [Read the crisis protocol ](https://safemeet.now/safety/crisis-protocol) report → count → ban ## Reports get counted. Bans get an expiry. Reports aren't reviewed by a person — they're counted. Enough reports from different people against the same user, each within 24 hours of the one before it, trigger an automatic ban that expires on its own. An appeal goes to an AI review: a ban it finds unjustified is lifted early, and otherwise the ban expires on schedule. [What happens when you report someone](https://safemeet.now/guides/what-happens-when-you-report-someone) Report filed User picks category + optional detail Reports counted Each within 24 hours of the last Threshold check Automatic, not read by a person Ban applied 24 hours, then it lifts itself Appeal reviewed By AI; an unjustified ban lifts early Coverage Every message Screened before delivery. No exceptions. Threshold Reports from 3 people Trigger an automatic ban, as long as each lands within 24 hours of the last. Fewer are needed when the reports say someone seems underage. Duration 24-hour ban Self-expiring. No manual unlock needed. Privacy Zero humans Nobody reads your messages to decide your case. child safety Child sexual abuse and exploitation get their own published standard: what is prohibited, how to report it from inside a conversation, how the material is handled, and the point of contact for it. [Read the child safety standards](https://safemeet.now/safety/child-safety-standards) safety guides Help that does not depend on SafeMeet: national crisis helplines, each checked against its own source, what to do if someone threatens to share your images, and guides to scams, bots, and chat apps for parents and teens. [Crisis helplines by country](https://safemeet.now/safety/crisis-resources)[What to do about sextortion](https://safemeet.now/safety/sextortion)[Scam red flags](https://safemeet.now/safety/scams)[How to tell if you are talking to a bot](https://safemeet.now/safety/spot-a-bot)[A parent's guide to anonymous chat apps](https://safemeet.now/safety/parents)[A guide for teens](https://safemeet.now/safety/teens) ready when you are ## Now you've read it. Go meet someone. [Start chatting ](https://app.safemeet.now/inbox)[How matching works](https://safemeet.now/how-it-works) --- Source: https://safemeet.now/safety