How Ava works
Ava is built to matter less in your life over time, not more. This page explains how, in plain terms, so you can check it against what the app actually does.
It is built to need you less
Ava compares how often you talk here with how often people in your life come up, and with the conversations you log with them. Each week it shows you that balance, compared only with your own previous week.
If you start leaning on it more while people come up less, it changes on purpose. Check-ins drop to 80%, then 50%, then 25% of their usual number. Replies get shorter, and it says so. It asks about the people in your life more often.
When check-ins shrink because things look steadier, it tells you that too. Fewer check-ins is the goal.
Check-ins are careful and rate-limited
Ava writes first only when at least two of four trend detectors agree that something has changed, or on an isolated evening. Never during your quiet hours, never while you are mid-conversation, and never beyond a small daily and weekly budget. A morning hello happens only if you turn it on.
A silence on its own is not a reason to write; needing this less is the point. It writes after a silence only if your recent pattern looked like withdrawal. If check-ins go unanswered or are marked not useful, it backs off. You can turn check-ins off at any time.
You can see why it replied the way it did
Every message gets a read on eight axes: joy, trust, fear, surprise, sadness, disgust, anger and anticipation, each from 0 to 1. A language model estimates them from your words. It is a guess, not a measurement, and the caption under each reply says how confident it is.
If the read is wrong, tap not quite and say what it was. The next reply owns up to the mistake, and later replies keep what you said in mind, the way they keep a memory. Reads are never adjusted automatically.
Your profile shows how the read has changed week by week.
You control what it remembers
When Ava notices something worth remembering, it asks first. Nothing is kept until you tap keep. You can edit any memory, or forget it with one tap. Forgetting deletes it for good.
Every memory is listed on your story page and in the Mirror. When a reply draws on one, it shows you which.
In a crisis, it puts real help on screen
If you write something that clearly signals suicidal thoughts or self-harm, verified helplines appear at once, before the model even replies. Softer signals ask first whether you want to see them. The Help now button is on every page, and the help page works without an account.
Find help near me asks for your location only when you tap it, uses it for one search, and does not store it. The hospitals it shows come from a list checked against each institution's own site; if you would rather not share your location, you can pick your state instead. Either way the helplines and a maps link stay on screen. Ava never makes up a number or a place.
The model does not try to talk anyone out of how they feel. Its job in a crisis is to stay present and help you reach a person who can help.
It will not pretend to be your person
Ava never says it is always here, that something can stay between you, or that it understands you better than anyone. A check runs on every reply before it is sent, and removes lines like these if the model writes them.
The same check removes probing questions about something painful you shared, labels like PTSD applied to you, and any phone number that is not on the verified list.
It does not rewrite itself
Ava does not change how it talks based on how you react: not your ratings, not how long you stay, not how you seem to feel after a reply. A system that tunes itself toward feeling good right now learns to flatter and stay available, which is the opposite of what this is for.
It improves another way. Ratings, corrections and check-in feedback are counted, without anyone's words, in a summary that people at Ava read. Any change they make is written by a person and must pass a test suite on honesty, crisis handling, memory and dependency language before it ships.
What it is not, and what is still being checked
Ava is software. It is not a therapist, a doctor, or a crisis service, and it cannot keep anyone safe. It is for adults, and for 13 to 17-year-olds with a parent or guardian's consent.
The crisis thresholds, the trauma-informed conversation rules and the language guard were reviewed by the team on 14 September 2026 against published guidance on safe messaging about suicide and trauma-informed care. A licensed clinician has not yet reviewed them; that review is planned before wider release.
Questions people ask first
Can anyone else read my conversations?
No one is shown them. They sit in a database in Mumbai that only Ava's own server can reach, and the part needed for each reply goes to the AI provider with your phone numbers, emails and ID numbers stripped out first. Nobody at Ava browses them: what people here read is a summary of counts, with nobody's words in it.
Will it tell my parents, my college or my employer?
No. It contacts nobody on your behalf and has no one to contact: it holds no phone numbers, and the only messages it writes are drafts that open in your own app, for you to send or delete.
What happens to it all if I stop?
Conversations are deleted after 30 days by default, and you can make that shorter or turn saving off entirely. Download everything or delete everything from your profile, whenever you like. Deleting is permanent and immediate.
Does it record me, or watch me?
Only if you switch a channel on, one at a time: voice tone, typing rhythm, the face channel. All three start off. The face channel never sends video anywhere - it runs on your device and sends two numbers.
Is it a therapist?
No, and it will tell you so when it matters. It cannot diagnose, cannot keep anyone safe, and is not a crisis service. In a crisis it puts real, verified helplines on screen before it writes anything.
What does it cost?
Nothing. No ads, no paid tier, nothing locked, and no data sold. Mental health support is a right, not a subscription.