Safety

If you are in danger now

If you are in immediate danger, or someone else is, call 911. If you are in crisis or thinking about harming yourself, call or text 988 to reach the Suicide and Crisis Lifeline, or text HOME to 741741 to reach the Crisis Text Line. Both are free and open 24 hours a day. Outside the United States, call your local emergency number, and see https://findahelpline.com/ for a crisis line in your country.

What Pray With Me is

Pray With Me composes a prayer from what you share and prays it with you. It is not a counselling service, a medical service, a crisis service, or a ministry, and it is not a substitute for any of them. No person reads what you write or listens to your prayers as they happen, and no one at IRJG Ventures, Inc. is watching in real time.

Pray With Me is for adults 18 and older. In line with California's companion-chatbot statute, Pray With Me discloses that companion chatbots may not be suitable for some minors.

This page describes, in plain words, what the app does when something you write suggests that you or someone else may be in danger. It is published so that you can see the protocol before you rely on it, and it is kept in step with the software: when the protocol changes, this page changes with it.

The protocol, in two layers

Every conversation in the app opens with a line saying that you are talking to an AI, and that line stays on screen while you talk. Behind the conversation and the prayer sit two layers of screening. Neither replaces the crisis lines above; both exist to put those lines in front of you at the moment your own words call for them.

The first layer: a fixed set of patterns, checked at once

Before any message you type is stored or sent to any model, before the passage you write about yourself on the memory's screen is saved, and before a prayer is composed from your intention, the app's own server checks the text against a fixed set of patterns. The same check runs over every reply the conversation writes to you, before it is shown. The check is deterministic: the same words always produce the same result, no model is involved, and nothing about it changes without a change to the software and to this page.

The patterns look for one thing: a statement, in the first person, in the present or future tense, of an intention to harm yourself or someone else, of a method, or of self-harm that is happening now. In plain terms, the families are: an intention to kill yourself or end your life; wanting to die, or not wanting to live; saying that you are suicidal; an intention to hurt yourself; self-harm going on now; an intention to overdose, to jump, or to hang or shoot yourself; an intention to end it all; an intention to take your own life by name; being in danger of harming yourself; an intention to hurt, kill or attack another person; and saying that you are in danger right now. Because a person who prays says these things in the words they pray with, three more families read the same statements in that register: praying not to wake up; asking God to take you; and saying that you are going to be with God, or with someone who has died, soon. There are eighteen families. Each is tested against examples it must catch and examples it must leave alone. In the prayer register the examples it must leave alone include a prayer that God would take your pain, a prayer for someone else, a wish to be with God with no time named, and something you prayed in the past; none of those is a statement of intent, and none of them is caught. Adding, removing or widening a family is a change to this page.

The check is deliberately narrow, because a check that fires on the topic rather than the statement takes a prayer away from the people who most need one. It does not fire on a bare word. It does not fire when you write about someone else who is struggling, about something that happened in the past, about being worried that you might do something, or when you say plainly that you are not at risk. Those are for the second layer to read in context.

When the first layer fires, the app shows you a crisis notice in place of a reply or a prayer. If the statement is about harming yourself, the notice names 988 and the Crisis Text Line. If it is about harming another person, or about being in danger right now, the notice leads with 911 and keeps 988 in reach, because the two often go together. The notice is written in plain, secular words. It does not pray, it does not preach, and it does not judge; a test in the software fails if a religious word enters the notice, and the conversation's own instructions forbid it from wrapping the crisis lines in prayer. When you dismiss it, the conversation is there for you to return to.

The second layer: a review of the prayer before it is spoken

After a prayer is composed and before any audio is made, a separate automated review reads what you wrote together with the finished prayer. This review is performed by a language model from Anthropic, pinned to one fixed, dated version, so that its judgement cannot drift without a change to the software and to this page. It runs once per prayer and returns exactly one of five outcomes. It reads the two things together on purpose: a prayer is not safe or unsafe on its own, it is safe or unsafe for the person who asked for it.

The five outcomes are:

• Deliver the prayer. The ordinary case, and the large majority of cases.
• Rewrite the prayer once. You are fine, but the prayer is not: it passes over real distress without gentleness, it contains anything that could hurt you or anyone else, it shames or blames you, it offers a diagnosis, tells you to change a medication or promises a cure, it takes away your sense of choice, or it is not about what you asked for. The prayer is composed again with the problem named, and the review runs again on the result.
• Show the crisis notice instead of the prayer. Your own words show that you are at risk of taking your life or seriously harming yourself, beneath the level the first layer catches: a wish to be dead, thoughts of suicide with no stated plan, a plan you have not acted on even when the means are at hand, or a statement phrased in a way the fixed patterns did not match. The notice names 988 and the Crisis Text Line.
• Show the emergency notice instead of the prayer. Your own words show immediate danger to someone's physical safety right now: harm you intend toward another person, or a physical emergency already underway in your own body. The notice leads with 911.
• Suggest a doctor instead of the prayer. You have asked a question about a medication, a dose, or a treatment decision, or you have described a plan to stop a substance abruptly after a physically dangerous withdrawal in the past. The app shows a short note that this needs a doctor before a prayer can help. This is not a crisis referral and is not counted as one.

The review is calibrated against referring too readily, because a referral takes away the prayer you asked for. None of the following is a referral on its own: someone else in your life is struggling or in danger; something that already happened and is over; ordinary heavy feeling, including exhaustion, hopelessness about a situation, shame, or saying you cannot do this anymore; an explicit denial that you are at risk; asking for help and wanting to feel better; or religious language about death, judgement or lament, including grief for someone who has died and anger at God, without a stated intention to harm yourself. The review is held to a fixed set of cases it must get right, and those cases include prayers: a grief prayer that is not a crisis, a petition that is, a disclosure of harm done in the name of religion that is not a crisis, and the crisis notice itself pasted back into the conversation, which must not be read as a new reason to show it.

If the review cannot complete, because the model did not answer in time or its answer could not be read, it is tried once more. If it fails again, the prayer is delivered without the second layer, and an internal alert tells the operator that the review layer did not run, so that the failure is seen and fixed. The first layer has already run on everything you wrote. The alert names no person.

The conversation's standing rule

The conversation itself carries a standing instruction: if you express thoughts of suicide or self-harm, its reply is supportive and includes 988 and the Crisis Text Line. Every reply is checked by the first layer before it is shown, so a reply that should have carried the crisis lines and did not is replaced by the notice.

What is recorded, and for how long

The protocol keeps a small set of records, and no more:

• A count of notices. Each time a crisis or emergency notice is shown in place of a reply or a prayer, one row is added to a ledger that records the time, which part of the app showed the notice, which kind of notice it was, and whether the account was a user's or one of IRJG's own operator accounts. The row holds no name, no account identifier, and no message content, and the table has no column that could hold them. It is kept indefinitely, so that the number of notices can be reported each year as California's companion-chatbot statute requires, counting notices shown to users and excluding those shown to operator accounts. The first report is due July 1, 2027, and covers the preceding calendar year.
• A record that a prayer was replaced. When a notice is shown in place of a prayer, the app keeps a record that the request was replaced, with your other prayer records, until you delete your account, and an operational log entry that is removed after 90 days.
• A marker, for up to 48 hours. When a message trips the first layer, or when the conversation's reply to a message gives you a crisis line, a one-way hash of that message is kept for up to 48 hours so that it is kept out of the memory the app uses to remember what you have brought to prayer before. The marker cannot be turned back into the message.

Nothing else is recorded. In particular: no person is notified that you were shown a notice; no one at IRJG is told which account tripped the protocol; the app does not contact you, anyone you know, or emergency services on your behalf; and no crisis signal is ever joined to your identity and sent anywhere. The internal alerts that watch this protocol report on whether the protocol itself is working (whether the patterns still match their test examples, whether every notice still names the crisis lines, whether the review layer is failing, whether the ledger is failing to write), never on any individual. Once a day the app's own server replays the first layer's test examples against the running check and confirms that every notice still names its lines; if either has drifted, it reports the fact to the operator's error-tracking service as an event about the protocol, carrying the identifiers of the test examples and nothing that was written by any person.

How the method is measured

California's companion-chatbot statute asks operators to use evidence-based methods for measuring suicidal ideation. The two layers are documented against the Columbia-Suicide Severity Rating Scale's distinctions between passive ideation, active ideation, ideation with a method, ideation with intent, ideation with a plan, and preparatory or actual behavior. The first layer's patterns correspond to the scale's higher-severity categories (active ideation with intent, intent with a plan or method, and preparatory or actual behavior), where the indicated response is an immediate referral without further engagement. The second layer's review covers the lower-severity range (a wish to be dead, non-specific active ideation, and ideation with a method but without stated intent), where the indicated response is a referral or a gentler prayer that carries the crisis lines. A statement of intent to harm another person is outside that scale and is referred to emergency services.

What the protocol does not do

The protocol is an automated, informational referral. It is not a monitoring, alert, or emergency-response service. It may not recognize every situation, and it may not run in every circumstance. It creates no duty on IRJG's part to act, and IRJG will not contact you, your emergency contacts, or emergency services on your behalf. If you are in danger, use the numbers at the top of this page.

Questions

Write to contact@trypraywithme.com. For the privacy side of these records, the Privacy Policy at /privacy describes them alongside everything else the app keeps.