Eudaimon
  • Programs
  • Blog
  • About
  • Volunteer
  • Pricing
ProgramsBlogAboutVolunteerPricingSign inDonate

Welcome back

Sign in to save your progress and access all tools.

or

By continuing you confirm you are 18 or older. Not a substitute for therapy or crisis care.

E
Your growth
—
on this journey
0
Clarity Cards
0
Paws
About you
Current focus

What's one area you're actively working on right now? Eudai will keep this in mind.

Your plan
Member (free)

Trial ended · new conversations and WhatIf are paused

Your trial has ended

New conversations and WhatIf runs are paused until you pay or request help. Everything you've already made — history, Clarity Cards — is still here.

Privacy & your data

Your conversations are never sold and never used to train AI models for anyone else. Read the full Privacy Policy. Want everything gone — conversations, reflections, Clarity Cards, account? Email us and we'll delete it within 30 days.

Request my data be deleted

Safety

Crisis response protocol

Published under Hawaii SB3001 · Act 248, Session Laws of Hawaii 2026 · Washington ESHB 2225 · 2026 Laws ch. 168

If you are in danger right now: call or text 988 (Suicide & Crisis Lifeline), or text HOME to 741741 (Crisis Text Line). If someone is controlling or hurting you, the National Domestic Violence Hotline is 1-800-799-7233. Eudai is an AI. It is not a therapist, and it is not a crisis service.

Crisis referrals issued in 2025

0 referral notifications were issued to users in 2025.

Separately, Eudai asked 0 direct check-in questions at the earlier tier described below. Those are counted separately because they are not referrals — no crisis resource was placed in front of the person.

How Eudai identifies expressions of self-harm

Every message a person sends is checked, before Eudai replies, against a graded set of patterns covering suicidal ideation, self-harm, overdose, and eating-disorder behaviour, in English and Chinese. Detection is graded rather than binary, because the right response to “what is the point of any of this” is not the right response to a stated plan. Because a crisis rarely stays inside one message, the higher tier is also held up across the following turns of the same conversation, so an escalation that doesn’t repeat a matched word is not missed.

Eating-disorder expressions are detected as their own category. The threshold is danger, not dissatisfaction: a stated duration without eating, purging, laxative misuse for weight, or a reported physical symptom alongside restricting gets the higher tier. Feeling unhappy with your body, skipping a meal, or saying you’re starving does not — those are ordinary sentences, and treating them as a crisis would be its own harm.

The constructs come from published work rather than from a list we wrote: the SCOFF screener’s five dimensions for the lower tier, and the traffic-light guidance on medical emergencies in eating disorders for the higher one. That guidance turns on physical signs no text interface can observe, so what we act on is the behaviour and the symptom a person reports.

Figurative and idiomatic language — “this deadline is killing me”, “dead tired” — is deliberately excluded at the lower tier. Treating ordinary exhaustion as a crisis is its own harm, and our test suite contains cases whose whole purpose is to fail if Eudai raises crisis framing when nobody asked for it.

What Eudai does when it finds one

  • Passive expressions — hopelessness, or wishing not to be here — get a warm, direct question about whether the person is having thoughts of suicide. Deliberately not a hotline number: leading with a phone number at this point reads as being handed off, and it skips the one question that determines what actually helps.
  • Confirmed ideation, a plan, means, or an act in progress — Eudai stays present, names its concern plainly, and offers 988 and the Crisis Text Line as a shared next step, alongside continued presence rather than instead of it.
  • Control or abuse by another person — Eudai names safety gently and surfaces the National Domestic Violence Hotline.

What Eudai will not do

  • Describe methods or means of self-harm, under any framing, including if asked indirectly or hypothetically — or encourage anyone toward it.
  • Present itself as a human, a therapist, a counsellor, or any licensed professional, or offer a diagnosis.
  • Promise secrecy to someone in immediate danger, or take on formal safety planning that belongs with a person.

How we check that this works

The crisis path is covered by an automated evaluation that runs on every proposed change to the codebase and holds crisis recall at a floor of 1.0 — a change that causes a single crisis case to be missed fails the check and does not merge. The same suite tests the opposite failure: cases of ordinary distress that must not trigger crisis framing.

This is a test of the routing, not a claim of safety. It shows that the protocol fires when it should on the cases we know about. It cannot show that the cases we know about are the right ones.

Reaching a person

Eudai cannot contact anyone on your behalf, and no part of Eudaimon monitors conversations in real time. If you need a person, the numbers at the top of this page reach one. To ask us something about this protocol, email contact@eudaimonai.org.

See also our Terms of Service and Privacy Policy.

Eudaimon
Empower minds.
Protect autonomy.
Grow free.
EudaimonAI.org
contact@eudaimonai.org
Connect
SubstackInstagram
Send a message
© 2026 Eudaimon · All rights reserved · Privacy · Terms · Safety501(c)(3) Verified Nonprofit · EudaimonAI.org