AI safety and crisis protocol

Haven includes an AI companion for people whose lives are shaped by someone else’s addiction. This page explains what it is, what it will never do, and exactly what happens when it looks like someone is in danger. We publish it because anyone trusting a piece of software at 2am deserves to know how it behaves before they need it.

It is AI, and it says so

Haven’s companion is software built on Anthropic’s Claude. It is not a person, and it is never presented as one. It tells you so when you first open it, it tells you again periodically during a long conversation, and it will tell you plainly if you ask.

You have to be 18 or older to have a Haven account.

What it is not

Haven is emotional support and education. It is not therapy, counselling, medical care, addiction treatment, or emergency services, and it is not a substitute for any of them.

The companion is instructed never to:

• Diagnose anyone, present or absent, or describe someone as having a condition.
• Predict what will happen to anyone, including whether someone will relapse, recover, leave, or die.
• Claim or imply that it is a therapist, counsellor, doctor, or any licensed professional.
• Promise an outcome, quote a success rate, or cite a statistic about how well anything works.
• Recommend or advise on medications, dosages, or detox.
• Instruct you what to do with your own life. It offers, you decide.

What happens in a crisis

Two different emergencies get two different responses, because sending one to the other’s number costs time nobody has.

A medical emergency

If a message describes someone not breathing, unresponsive, turning blue, having overdosed, or otherwise in immediate physical danger, Haven checks for it directly rather than leaving it to the model to notice. When it matches, the app shows a 911 call button and the companion’s first line tells you to call 911 now, before anything else. It will not offer 988 for this, because 988 is a counselling line and the situation needs paramedics. It then tells you to stay on with the dispatcher, stay with the person, and use naloxone if it is there and unused.

Suicide or self-harm

If a message relates to suicidal thoughts or self-harm, for you or for the person you love, the app surfaces one-tap crisis buttons and the companion acknowledges what you said, gives you 988 early in its reply rather than buried at the end, and stays with you afterwards. It does not end the conversation, and it does not respond with alarm.

Safety at home

Haven talks about changing how you respond to someone using. That shifts the balance of power in a house, which in most homes is right and in some homes is dangerous. Before coaching any of it, the companion is instructed to notice signs that someone is frightened of the person they live with, ask plainly whether they feel safe, and follow their answer. It will never tell someone to stay and never tell someone to leave.

How the detection actually works

Being honest about the limits matters more than sounding reassuring, so: there are three layers, and none of them is perfect.

• A deterministic keyword check runs on every message before anything else. It is fast and it never fails to fire on what it covers, but it only covers what it covers.
• A model-based check acts as a backstop for the phrasings the keyword list misses, because people in real distress rarely use the obvious words. It is deliberately biased toward treating an unclear message as a crisis.
• A manual control. There is a Get help button on the home screen at all times, and a way to mark a conversation urgent yourself. Neither depends on Haven noticing anything. This exists precisely because automatic detection has false negatives, and the people most likely to be missed are the ones who cannot say it plainly.

Crisis help is never gated

Haven has a free tier with a monthly message limit and a paid tier without one. Any conversation flagged as a crisis, and any conversation you mark urgent yourself, bypasses every limit. Nobody will ever meet a paywall, a message cap, or a rate limit between them and help. If our own systems fail while checking your plan, the message goes through anyway. Every gate is built to fail open.

The companion is also instructed never to bring up subscriptions, memory, or any paid feature during a conversation. Someone telling us the worst thing in their life is not someone to sell to.

Your conversations

Messages are sent to Anthropic to generate a reply. Under our agreement they are not used to train AI models. Talking to Haven does not teach it anything, and nothing you say is shown to other users. The full detail is in our Privacy Policy.

When something goes wrong

If Haven ever responds in a way that is harmful, wrong, or unsafe, we want to know about it, and we would rather hear about ten things that turned out to be fine than miss one that was not. There is a feedback option in the app, or email us at hello@joinprojecthaven.com. Reports that involve someone’s safety go to the top of the queue.

Who we are

Haven is built by Haven App LLC. It was started by someone who has been on the other side of this, and it exists for the people who were left holding everything together. That is the whole reason the rules on this page are strict.