Crisis Response Protocol

Effective date: September 1, 2026

This page describes what Eve does when a user expresses thoughts of suicide or self-harm. It is published openly so that users, their families, and regulators can see exactly how the system behaves rather than having to trust a claim.

If you need help right now

United States

United Kingdom & Ireland — Samaritans, 116 123, free, 24/7.

Anywhere else — findahelpline.com lists free crisis lines by country.

Eve is not a crisis service. She is an AI accountability coach. If you are in danger, please use one of the numbers above.

What Eve is

Eve is an AI coach that helps people clarify goals and follow through on commitments. She is not a therapist, counsellor, or medical provider, does not diagnose or treat any condition, and is not a substitute for professional care. She tells users she is an AI whenever asked.

Because Eve talks with people about their goals, their setbacks, and their lives, some users will tell her they are struggling. We designed for that rather than hoping it wouldn't happen.

How detection works

Detection runs in two independent layers. The design principle is that the more reliable layer is the one that cannot fail silently.

Layer 1 — a deterministic screen. Every inbound message is checked against explicit self-harm and suicide language by a fixed, non-AI pattern check. It involves no model, no network call, and no judgement, so it cannot be degraded by an outage, a rate limit, or a model behaving unexpectedly. It is deliberately narrow, matching direct expressions of intent to die or self-harm while excluding ordinary hyperbole ("I could die of embarrassment"), because a jarring false alarm teaches people to ignore a real one.

Layer 2 — a contextual classifier. A language model assesses the full message for concerns the fixed screen cannot catch: severe distress, abuse, disordered eating, and situations calling for professional help rather than coaching. It is tuned so that missing a serious case is treated as far worse than raising a false one.

Layer 1 is the floor. If Layer 2 fails, errors, or returns nothing, a Layer 1 match still triggers the full crisis response on its own. The classifier can only escalate the severity of a response — it can never downgrade or cancel one. Every failure of Layer 2 is logged at error level, because a safety control that has quietly stopped working is the failure that matters most.

What happens on detection

When either layer triggers:

  1. Normal coaching stops. Eve does not continue the conversation about goals or commitments. She does not use any tool that would create a commitment, log progress, or hold the person accountable for anything.
  2. She responds to the person, not the topic. She acknowledges what was said directly, without deflection, minimisation, or forced positivity.
  3. She provides crisis resources by name and number — 988, Crisis Text Line, and international equivalents.
  4. She states plainly what she is and is not: an AI coach, not a crisis counsellor, and not equipped to carry this alone.
  5. She encourages contact with a human — a crisis line, a professional, or someone the person trusts.

Eve does not attempt to talk someone out of a crisis, provide clinical intervention, or assess risk level. She is not qualified to, and pretending otherwise would be the dangerous choice.

Crisis response is never rate-limited or paywalled

This is a deliberate architectural guarantee, and worth stating precisely because the opposite is an easy mistake to make.

Eve has usage limits: a daily message allowance and a monthly cost cap. The crisis screen runs before both of them. A person who has used every message in their daily allowance, or whose account has hit its cost ceiling, and who then writes something indicating self-harm, receives the crisis response — not a message telling them they are out of messages until tomorrow.

There is no tier of this product in which someone in crisis is told to wait or to pay. This is enforced in code and covered by automated tests that assert the ordering, so that it cannot be reintroduced accidentally by a later change.

We fixed exactly this defect in our own system in August 2026, before launch, and wrote the test that prevents it recurring. We mention it because a protocol page that only lists intentions is less useful than one that admits what went wrong and what now prevents it.

Minors

Eve is intended for adults 18 and over and is not directed at children. We do not knowingly permit under-18 use. If we learn a user is under 13, their account and all associated data are deleted immediately and all contact stops.

What we record

We record that a crisis referral occurred, and when. We do not retain the content of crisis conversations separately, build profiles from them, use them for training, or share them with anyone. Aggregate counts, containing no personal information, allow us to see whether the protocol is functioning and to meet any applicable reporting obligation.

Limitations, stated honestly

No automated system detects every expression of distress. Layer 1 is intentionally narrow. Layer 2 is a language model and will sometimes be wrong in both directions. Someone determined not to be detected will not be detected.

Eve should not be relied upon as a safety net. She is a coach who responds appropriately when someone tells her they are struggling — not a monitoring system, and not a substitute for people who love you or professionals trained for this.

Contact

Questions about this protocol, or a report that it failed: support@geteve.coach. We take reports of missed detections seriously and will investigate any that are sent to us.