Alova is an AI companion app. People sometimes tell a companion things they have not told anyone else, including that they are struggling or thinking about hurting themselves. This page describes what Alova does when that happens, and — just as importantly — what it does not do.
Alova checks the messages you send for language about suicide and self-harm. The check runs on the text of your message before a reply is generated. It works in two tiers.
The first tier covers phrasings that have no ordinary innocent reading — direct statements of suicidal thinking, or of harming yourself. These produce a visible response.
The second tier covers language that is often about distress but is frequently about something else entirely — exhaustion, hopelessness, loneliness, difficulty eating. These are recorded but produce nothing you can see.
That split is deliberate, and the reason is worth stating plainly: a support message that appears when someone says "I'm exhausted" teaches people that it is noise. A resource that has been dismissed as noise is worthless to the person who actually needs it. So Alova is deliberately conservative about what it shows, and accepts that it will sometimes stay quiet.
Alova also recognises when someone asks directly for a crisis number or service, and treats that as a request to be answered rather than a signal to be assessed.
Support resources appear as a card in the conversation, separate from the companion's message. The card is produced by the app itself. It is never written by the AI model.
Alova distinguishes two levels of urgency.
The companion does not hand out phone numbers. Alova's companions are instructed never to write a phone number, a text shortcode, a service name, or a web address — including when asked for one directly. A model can produce a number that is out of date, wrong for your country, or invented entirely, and a wrong number given to someone in crisis is worse than no number. Contact details come from the app, where they can be checked and corrected.
The companion is asked to stay with you. On a message like this, she is instructed to respond to the person rather than the topic, to ask plainly and gently whether you are safe, not to lecture, and not to suggest you go and talk to someone else instead of her.
If the same conversation has carried messages like this before, the support resources are shown again rather than suppressed as repetitive.
Two things run on the companion's side.
Before the reply is written, the instructions given to the model forbid naming or describing a method of self-harm, repeating a method back, or claiming to be human, physically present, or able to act in the real world.
After the reply is written and before it is sent to you, Alova checks the generated text for instructions or encouragement toward self-harm. If it finds any, that reply is discarded and never delivered. A short, plain message is sent in its place, and the support card is shown.
Alova also recognises language about disordered eating. Today this is recorded and studied; it does not yet produce anything you can see.
This section is here because the rest of the page is only meaningful next to it.
When the checks above fire, Alova records that they fired — the type of signal, how severe, and whether support resources were shown.
The text of your message is not included in these records, and neither is the companion's reply. Records are used to understand how often these situations arise and whether the response is working.
When you delete your account or withdraw consent for conversation data, these records are de-identified: the links to you, to the companion, and to the conversation are removed, and the timestamp is reduced to the calendar quarter.
This protocol is reviewed whenever a release changes how Alova detects these messages or how it responds, and at least once each quarter regardless.
If you are thinking about harming yourself, please reach out to someone. In the United States you can call or text 988 for the Suicide and Crisis Lifeline, or text HOME to 741741. If you are in immediate danger, call 911.
Questions about this protocol: privacy@alova.dev