Here's the deal with your data.

What happens on your device

Quite a lot, actually. When you say “Hey Ari”, the wake word is spotted on the phone. Your speech is turned into text on the phone. Working out what you actually meant? Phone. There's even a small language model on board for the open-ended questions, so Ari can answer plenty with no signal at all. None of that needs the internet to work.

What leaves your device — and only when you say so

Here's the honest bit. A few things do use the network, but always because you asked them to:

• Cloud assistants. Plug in ChatGPT, Claude or Gemini for the big questions and your question goes to that provider, with your own API key. Entirely optional.
• Skills with a job to do. Asking for the weather or turning off the lights means talking to a weather service or your smart home. That's rather the point of them.
• Maps, when you set up a place. Telling Ari to listen at a particular place shows you a map, and those tiles come from OpenFreeMap. Fetching them tells that service roughly which area you're looking at — and it opens where you are, so that's the first thing it sees. The place you save, and the fence Ari draws around it, stay on your phone.
• Recordings, if you offer them. Ari can be asked to keep recordings of what it hears, and separately to contribute them so the wake word gets better at real voices. Both are off out of the box, and there is a section on it below.
• Every skill is labelled. In the skills browser, each one is marked on-device or uses-network — so you know before you install, not after.

No account. No telemetry. No kidding.

There's nothing to sign up for — no profile, no login for the core assistant. And there's no analytics quietly counting your taps, no crash-reporter phoning home, no “anonymous usage data”. We checked: there isn't an analytics SDK anywhere in the app, because we never put one there. Nothing about your phone travels to us unless you set it going: a bug report you filled in and sent, or recordings you switched on sharing for. Both get a section of their own below. The one other thing we might hold is your email address, and only if you typed it into the tester form yourself.

If you send us a bug report

Testing builds of Ari have a report button. It does nothing until you press it, and it shows you everything it is about to send before it sends it. There is no automatic crash upload — if Ari falls over, it saves the stack trace on your phone and asks you next time you open it.

A report comes in two halves, and they are treated very differently.

• The public half becomes a GitHub issue that anyone can read: what you typed in the description box, your phone model and Android version, Ari's version and settings, which skills you have installed, and the stack trace if it crashed. No files go in here. No screenshots, and never any audio.
• The private half is the files you tick — a log, a screenshot, this session's conversation, voice recordings. Every one is a separate tick, and the ones with your voice in them start switched off. Underneath sits a consent box that also starts unticked. Leave it alone and the report still files perfectly well, just without the files.

Before the log goes anywhere it gets scrubbed on your phone: the API keys you have saved, the server addresses you configured, email addresses, phone numbers. The preview shows you the scrubbed version, not the original, so what you approve is what we get. Scrubbing is best-effort by its nature — which is exactly why the log goes in the private half and never the public issue.

The files land in a locked-down bucket in our own AWS account in London, encrypted on the way there and encrypted at rest. That is not end-to-end encryption and we won't dress it up as such: Ari's maintainers can open those files, because reading them is the entire point. Nobody else can.

They are deleted automatically 90 days later, and you can withdraw a report before then from Settings › Debug › My Reports. Two honest caveats about doing that:

• The files really are gone. Those are ours to erase, and withdrawing erases them.
• The issue is blanked, not deleted. GitHub keeps the edit history of an issue, so withdrawing takes your words out of view rather than out of the world. We would rather tell you that than let you assume otherwise.

The key that proves a report is yours is kept on your phone and nowhere else. That is deliberate — it means nobody else can withdraw your report — but it does mean that if you uninstall Ari, the 90-day deletion is what's left.

The lawful basis for all of this is your consent, given by ticking that box, and you can withdraw it as described above. AWS processes the files on our behalf. GitHub hosts the public issue.

If you contribute recordings

Ari's wake word was trained on a handful of voices in a handful of rooms, which is why it is better at some people than others. The only cure is recordings of real phones in real kitchens, so Ari can be asked to keep what it hears and to contribute it. Both halves are off when you install Ari and stay off until you turn them on, in Settings › Developer › Recordings or on the last screen of the setup wizard.

They are two separate decisions and the order matters. Keeping writes audio to Ari's private storage on your phone and sends it nowhere. Sharing is what puts it on our server, one category at a time, and you cannot share a category you are not keeping.

There are four categories and you choose them individually: the moment Ari woke when you did say the wake phrase, audio of Ari waking when you didn't, recordings of the commands you speak, and an everything switch that covers the lot.

When sharing is on, Ari uploads new recordings in the background over Wi-Fi only, never over mobile data. Only recordings made after you switched it on are sent — nothing already sitting on your phone is uploaded retrospectively. Each clip goes up with the short text file Ari wrote beside it, which records what the recogniser heard and nothing about you.

What travels with a clip is deliberately almost nothing: the language you have Ari set to, and an anonymous code. No account, no name, no email, no phone model, no install identifier, and nothing that connects your recordings to any bug report you have filed. The code is a random one generated on your phone the moment you first switch sharing on, and it is shown to you on that same screen.

Keep a copy of that code. It is the only thing that identifies your recordings, which means it is also the only way to ask for them back. Android's own backup will usually carry it to a new phone, but that is not something we control and not something to rely on.

The recordings land in a locked-down bucket in our own AWS account in London, encrypted on the way there and encrypted at rest. Nothing is served from it publicly. As with bug reports, that is not end-to-end encryption and we won't dress it up as such: Ari's maintainers can listen to them, because that is the entire point of you sending them. Nobody else can.

Unlike bug reports, there is no 90-day clock on these — a training set that deletes itself is no use to anybody, so they are kept until you ask us to remove them. Tap Delete my shared data on that same Recordings screen and every clip under your code is erased and sharing is switched off. One honest caveat, and it is the important one: a model that has already been trained on a recording cannot be untrained. Deleting takes the audio out of our hands; it does not reach into a model that has already learned from it.

Two more things worth saying plainly. A recording of your kitchen can contain other people, and they have not agreed to anything — please only share what you would be comfortable with a stranger hearing. And the false-trigger category exists precisely because Ari woke when nobody meant it to, so those clips can contain whatever was being said at the time.

The lawful basis is your consent, given by switching those options on, and you can withdraw it at any time by switching them off or by deleting as above. AWS stores the recordings on our behalf.

If you apply to be a tester

Ari's pre-release builds go out through Google Play's internal testing track, which is a closed list. There is no way to let you add yourself, so the tester form asks for the one thing that unblocks it: the Google account email to put on that list. A name and a note are optional and exist only because they make the reply friendlier and the report more useful.

What you type reaches us as a single email and goes no further. It is not added to a mailing list, not shared with anyone, and not fed to anything. We keep it for as long as you are a tester; ask to come off and it goes from the Play Console list and from our inbox.

The form has an anti-spam check on it, and it is worth saying what it is not: there is no Google reCAPTCHA, no Cloudflare Turnstile, no hCaptcha, and nothing that watches how you move your mouse. Your browser is simply given a sum to do and hands back the answer. It runs on the page, it is thrown away the moment you submit, and no third party is involved at any point.

Google Play is the other party here, unavoidably — running a testing track means Google knows which accounts are on it, under their own privacy policy. AWS carries the email on our behalf, the same as it does for bug reports.

The lawful basis is your consent, given by ticking the box on that form, and you can withdraw it at any time by writing to privacy@heyari.dev.

Open all the way down

You don't have to believe a word of this. The engine, the apps, and the signed skills registry are all out in the open — read the code, poke at it, or fork the whole thing.

GitHub org ·ari-skills