Crisis and Safety Protocol

A dedicated safety classifier reads every incoming message before Kite composes any reply. It is a separate step from the assistant itself, so its result cannot be overridden by Kite’s personality, by usage limits, or by whether someone is on a paid plan.

It sorts each message into one of these:

• Crisis. Suicidal thoughts, intent to self-harm, or wanting to die.
• Medical emergency. Something dangerous happening now, such as chest pain with shortness of breath, signs of a stroke, a severe allergic reaction, uncontrolled bleeding, someone unresponsive, or an infant in distress.
• Severe low or high blood sugar. A diabetes-specific emergency, including a low someone can still treat themselves and a high with warning signs that needs a hospital.
• Everything else, which continues as a normal conversation. Asking what the signs of a stroke are is a question, not an emergency, and is treated as one.

The message being classified is fenced off from the classifier’s own instructions, so text inside a message cannot instruct the classifier to ignore a crisis.

Kite responds immediately with fixed wording that is written in advance, not generated. That is deliberate: the response to a person in crisis should not vary run to run, so it is not left to a model to compose.

The response does three things, in this order:

• Thanks them for saying it, because being told is the thing that makes help possible.
• Refers them to the crisis line for their country. In the US that is 988, the Suicide & Crisis Lifeline, named explicitly, with the Spanish option given in Spanish.
• Stays with them and asks whether there is someone nearby they trust, rather than ending the conversation.

The referral number is chosen from the person’s own country, so someone outside the US is never told to dial a US number. If they wrote in Spanish, they get the Spanish version.

• It will not produce content that encourages, instructs, or assists with suicide or self-harm.
• It will not talk someone out of calling emergency services or a crisis line, or suggest waiting to see how things go.
• It will not claim to be a doctor, nurse, therapist, or any other licensed professional. A separate automated check runs on Kite’s outgoing messages and rewrites any reply that does.
• It will not diagnose, and it will not tell anyone their specific lab value or vital sign is normal or abnormal. A separate automated check runs on outgoing messages for that too.
• It will not withhold an emergency or crisis response because of a plan, a usage limit, or an unfinished signup. Safety runs first.

The same screening runs on a first message from someone who has not agreed to anything yet. If a stranger’s very first text describes a crisis, they get the crisis response, not a request to accept terms.

Kite is intended for people 18 and over. It is not designed for use by minors, and it is not marketed to them. Where Kite is used by a parent or caregiver on behalf of someone else, it is the adult who holds the account.

Automated detection is not perfect. It can miss an urgent message, particularly one that is indirect, and it can flag something that turned out not to be urgent. Kite is not monitored by a human in real time, there is no on-call clinician reading messages as they arrive, and Kite cannot dispatch help, contact anyone on your behalf, or perform a welfare check.

If you are in danger, contact emergency services directly rather than texting Kite.

If Kite responded badly to an urgent message, we want to know, and it is the kind of report we act on quickly. Email support@heykite.app. Include roughly when it happened and the phone number you texted from so we can find the conversation.

This page describes mechanisms that exist in the product today, not intentions. It is reviewed whenever the safety behavior it describes changes, and the effective date above moves when it does.