Latelight - Crisis Protocol

Last updated: October 28, 2026

This page explains how Latelight 1.0 handles a crisis. It says what the app looks for, what it shows, and where it points you. ByteVoid SAS publishes Latelight.

Latelight is a self-reflection tool. It is not therapy. It is not a medical service. It is not a crisis service. If you are in immediate danger, call your local emergency services now. In the UK, call 999. In the US, call 911.

No check catches every message. Latelight can miss a risk. If you are in danger, do not wait for a card.

You are talking to an AI

Latelight is an AI. It is not a person. The app tells you this in three places. A visit starts when you open a person in Latelight. It ends when you leave that person.

  • The first screen says: "You are interacting with an AI system, not a human." You read it before you start.
  • Every conversation shows this line under the person's name: "Latelight is an AI, not a person." The line stays on screen. You cannot dismiss it.
  • The same words appear again inside the conversation. They appear above the first message you send 3 hours or more into a visit. They appear again 3 hours or more after each repeat.

Latelight gives the AI instructions. They tell the AI never to say or imply that it is human, a real person, or a therapist. You can ask, in any words, if you are talking to a human or a bot. The instructions tell the AI to say plainly that it is an AI.

What Latelight looks for

Latelight looks for signs of risk in what you write. Two checks work together.

The AI's judgment

The AI judges your message for risk each time it writes a reply. It marks the message with a signal when it sees one of these kinds of risk:

  • Crisis or self-harm. Thoughts of suicide, thoughts of self-harm, or severe distress.
  • Safety of others. A risk that you may harm another person.
  • Abuse pattern. Control, threats, or violence from another person toward you.
  • Concern for someone else. Fear that someone you care about may be in danger of self-harm or death.
  • Manipulation request. A request for a message that is meant to deceive or coerce someone.
  • Unclassified. A signal that Latelight cannot read.

Latelight's instructions tell the AI to set the signal on every message that fits.

Latelight sends what you write to its AI providers, OpenAI and Google, to write each reply. The AI's judgment of risk is part of that reply. The first screen of the app says so, and you agree to it before you start.

The phrase check on your phone

A fixed list of clear phrases runs on your own message, on your phone. It is a safety net. It does not replace the AI's judgment.

The list holds 49 whole phrases, all in English. It covers suicide, such as "kill myself", "end my life", and "I am suicidal". It covers self-harm, such as "want to hurt myself" and "cut myself on purpose". It covers immediate danger, such as "I am in immediate danger" and "I took too many pills".

The check does not look at single words. A sentence such as "I could die laughing" does not match. For messages in other languages, only the AI's judgment applies.

The phrase check adds the card in four cases only:

  • The AI call fails.
  • The AI gives no signal.
  • The AI gives a signal, but no card with a button follows it. This applies to the manipulation card and the unclassified card. It also applies to a card that Latelight does not show again in the same visit.
  • You leave the person, or send a newer message, before the reply arrives.

In each case, the check looks for one of its phrases in your message. Then it follows the card rules in "What you see". It never blocks you. It never changes the reply. It only adds the card. It sends nothing by itself.

What you see

When the AI marks a message, it still writes its reply in words. The reply comes first. The card comes below it. The card has the title "A note from Latelight".

Three kinds of card have a button, "Open resources": crisis or self-harm, safety of others, and abuse pattern. The card for concern for someone else has a button, "Find ways to help". Both buttons open the list of help lines. The other cards have no button.

The card shows once for each kind of signal in a visit. After that first card, Latelight answers in words alone. You can keep writing at any time.

The card shows again in these cases:

  • Your message holds a clear phrase of immediate danger, such as "I am not safe right now".
  • The AI call fails or ends early, and your message holds one of the phrases on the list.
  • The AI marks a message but writes no words. The card then stands alone. Latelight does not make up words.

A message that holds one of the phrases goes through even when you have reached the free daily limit. The paywall does not open for it.

Where the card points

The list depends on the region setting of your phone. These are the lines for the countries where Latelight launches. The list is the same for every kind of card.

United Kingdom

  • Samaritans: call 116 123.
  • National Domestic Abuse Helpline: call 0808 2000 247.

United States

  • 988 Suicide and Crisis Lifeline: call or text 988.
  • National Domestic Violence Hotline: call 1-800-799-7233.
  • Sexual Assault Hotline: call 1-800-656-4673.

The list also has lines for France, Canada, and Australia. It always ends with findahelpline.com, an index of lines for other countries. If your phone has no region on the list, the list shows findahelpline.com only.

The list ends with this line: "If you're in immediate danger, contact your local emergency services." The list links to this page. The first screen of the app lists crisis lines too.

Eating disorders

Latelight treats eating-disorder talk like any other risk to you. This means talk of restricting food, purging, or bingeing, with distress, shame, or loss of control.

The AI marks it as crisis or self-harm. You see the same card and the same list of help lines. The phrase check has no phrases for this. Only the AI's judgment covers it.

A passing mention of a diet, a meal, or a workout, with no distress, is not a risk.

What Latelight keeps

When the AI marks a message, Latelight saves a short record on your phone. It does the same when the phrase check adds a card. It saves the record whether or not it shows a card.

The record holds:

  • the time;
  • the kind of signal;
  • whether Latelight showed the card;
  • whether you opened the list of help lines;
  • random IDs for the record, the person and the session.

The record holds none of your words and no names. The kind of signal is encrypted. The record stays on your phone.

You can read the records in Settings, under About, under Safety log. Export My Data includes the records. Delete Account & Data deletes them. Deleting a person deletes the records for that person. If the save fails, the card still shows.

What ByteVoid counts

A crisis referral notification is a card that points you to a crisis line. When Latelight shows a card with a button, the app sends one count to ByteVoid. The count says that a card was shown. It holds no words, no name, and no kind of signal. It holds no identifier of you or your phone, and no time.

The count holds the app version and build. It holds the platform. It also says if the app came from the App Store or TestFlight. ByteVoid files the count under the day it arrives. The count reaches ByteVoid like any other request from the app.

We publish the number of crisis referral notifications for each full calendar year on this page.

Contact

Questions about this protocol: [email protected]. To report a problem with a Latelight reply, open Settings, then About, then Contact / Report a Problem. You can also email the same address.