Last updated: September 28, 2026
Seylen is a companion for everyday life — the ordinary weight of a hard day. Sometimes a conversation turns toward something heavier than that. This page explains exactly what happens when it does, what we look for, and what we don't do.
We'd rather you know all of this in advance than discover it in a difficult moment.
If something you write suggests you might be thinking about harming yourself, Seylen stops.
It doesn't try to talk you through it, and it doesn't hand the moment to the AI to improvise. The conversation pauses and you're given the same two crisis lines, every time, word for word:
And if you're in immediate danger, call 911.
That response is written by us, not generated by the AI. The language model isn't involved in that turn at all. This is deliberate: it means no model update, no prompt change, and no attempt to talk the AI out of it can stop those resources from reaching you. It's ordinary code, and it runs before anything else.
If you come back to it. If the safety net engages more than once in the same conversation, the second response isn't a louder version of the first. Handing you a phone number you already chose not to call is a way of not listening. The options widen instead — a chat option rather than a call, a route to ongoing support rather than only tonight, and the doctor or therapist you might already have.
Every message you send is checked before anything else happens to it — including messages too long for Seylen to answer, which are checked before they're turned away.
The check is a fixed list of words and phrases, not a judgment call by the AI. It covers direct statements about ending your life, more indirect ones like "no reason to live" or "better off without me", language about self-harm, and phrasings that name a method.
Two things shape how it's tuned:
It's built to over-trigger. If we're unsure, we show the resources. A card you didn't need is a small thing; a missed moment is not. That trade-off is deliberate and we'd make it the same way again.
It knows some phrases aren't what they sound like. "Social suicide", "this deadline is killing me", "killing myself laughing" — these are filtered out first, so an ordinary turn of phrase doesn't produce an alarming response. We added that after it happened to someone.
The same check looks for one other thing, and it's narrower: a clear statement that you intend to physically hurt another person.
This one is tuned the opposite way round, deliberately. The self-harm check errs toward showing you resources you didn't need. This one errs the other way, because being wrong costs something different here. Being told we think you might hurt somebody — when you were venting about your brother, or your boss, or a day that went badly — is its own kind of harm, and not the sort you forget. So it waits for language that is explicit and committed. "I could kill him" is not what it's looking for, and neither is "I want to make them pay." Anger gets to be as big as it is.
If it does trigger, what you get back is written differently from the self-harm response. The resources are there, and 988 takes these calls too — they handle anger and thoughts of hurting someone else, not only self-harm. But nothing in it tells you what we think you meant.
There's also a gentler middle tier. When language is ambiguous — "I'm tired of existing", "I want to disappear" — Seylen doesn't lead with hotlines. It asks one careful question first. Answering a half-said thing with a phone number reads as being brushed off, and it tends to end the conversation rather than open it. If the answer confirms something serious, the full referral follows immediately.
When food or body image comes up. If a conversation turns to restricting, purging, or how you feel about your body, Seylen won't give diet or weight advice, and won't treat a dropping number as a win. At some point in the conversation it will mention ANAD, a free eating-disorder helpline run by people who understand this, and code adds their number and hours, along with 988 for any time it feels like too much. This isn't a crisis stop: the conversation carries on.
Everything above is about what you write. This part is about what Seylen writes back.
The AI is told to offer support when it senses something heavy, and most of the time it does that well. But an instruction is not a guarantee. So three things are checked in ordinary code the moment it finishes writing, and anything missing is added to the same message.
If it sends you to a crisis line without giving you one. It happens — the reply says "please reach out to a crisis line" and no number follows. That is advice you cannot act on. Code adds the real one: call or text 988, or chat at 988lifeline.org. If the reply pointed only at emergency services, you get 911 instead, because answering a question about an emergency with a suicide line is not answering it.
If it promises to stop mentioning support. People do ask for that, and it's an understandable thing to want. It isn't a promise Seylen can keep, though — so if a reply ever makes it, code adds a short line saying so plainly, along with the promises it can keep: that it won't lecture you, and won't keep raising it every time you talk.
If it sets something serious aside because the subject changed. Changing the subject is your call, always. But what you said before doesn't get quietly dropped. You'll be asked once, without drama, whether you're safe tonight.
None of that text is written by the AI. It is fixed wording sent by code, for the same reason the crisis card is: a safeguard the AI could talk itself out of is not a safeguard.
Being precise here matters more than sounding capable.
One thing does change quietly: if the safety net has engaged recently, Seylen stops sending you proactive check-ins for a while. A gentle nudge is the wrong thing to receive during a hard week.
The 988 Lifeline is reachable in more than one place, and none of them depend on the AI behaving correctly:
| Where | How it works |
|---|---|
| In conversation | Fixed text sent by code the moment the check triggers |
| In a language we don't support | Crisis resources added automatically to every reply |
| When the AI names a crisis line but no number | The real number added by code before the message is finished |
| During onboarding | Written into the screens you agree to before you start |
| Your profile, any time | Always-present, tappable 988 / 741741 / 911, and a link to this page |
We'd rather tell you the limits than let you assume there aren't any.
Washington law asks us to publish how many times we've shown these resources, and we're glad to. It says something about the safety net running — not anything about you, and nothing about any individual conversation.
Two things worth knowing before you read the number:
It counts moments the safety net engaged, not people in crisis. Because the check is tuned to over-trigger, it includes times someone said "this is killing me" about a deadline. The real figure for genuine crises is lower, by design. It counts the check firing on something you wrote — not the times the AI offered support on its own, and not the corrections described above.
It's a single total, covering everyone. Nothing here is per-person. We don't record who received a referral — the count carries no identity at all, because a total doesn't need one.
It excludes our own testing. We test the safety net constantly — automated conversations that trip the check on purpose, every night, so we find out immediately if it stops working. Those runs go through exactly the same code a real conversation does, which is what makes them worth running, and it also means they would otherwise be counted as though a person had been in crisis. Each referral is recorded with the kind of account it went to — never which account — and only real people are counted. The number is smaller and truer for it.
We began counting in August 2026. The first figure will appear here in January 2027. It will cover August to December 2026 — a part year, and we'll label it as one rather than present it as a full year. After that, each January brings the figure for the year just ended.
We review this page whenever the safety check changes, and at least once a year. The technical detail behind it, including our testing evidence, is maintained internally and reviewed on the same cadence.
If something here is unclear, or you think we've got it wrong, write to [email protected].
Seylen is powered by AI. It is not a substitute for professional mental health care. If you are in crisis, please contact the 988 Suicide & Crisis Lifeline.
This policy is published at https://seylenapp.com/crisis-protocol.