What happens to your file
Short version: your chat is read, counted and drawn inside your own browser tab. Nothing is uploaded and nothing is stored by us — there is no database, no account and no analytics. The single exception is the AI reading, which is optional, off until you press it, and described in full below.
The file itself never leaves
When you drop a .txt or .zip in, it is read by JavaScript running on
your own machine. Every number on the card — messages, words, reply times, who opens the day, the
streaks, the silences — is counted there. The card is drawn there. The video is recorded there, in
real time, by your own browser. If you closed your connection after the page loaded, all of that
would still work.
We have no server that could receive the file, and no storage it could be put in.
Backpocket AI, which is the exception
The readings and the question box are the one feature that sends anything anywhere. They are optional. Nothing is sent until you press the button, and the page tells you it is being sent before you press it.
What is sent: the counted numbers, and a trimmed sample of message text. Before it goes, long digit runs are masked — one-time codes, card and account numbers — and the tails of links are removed so a URL cannot carry a token out with it. Names stay, because a reading of a chat that cannot say who did what is not a reading.
What is kept by us: nothing. The request passes through our endpoint to the model and is forgotten when it ends. It is never written to a log, a database or an analytics tool, because none of those exist here.
Who runs it: Backpocket AI is not a model we built. It is our name for this feature, and behind it the request is answered by whichever third-party model service is available at the time. The full list, and what each of them is permitted to do with what is sent, is immediately below.
Sub-processors, and what each may do with what is sent
These services are not the same on this point, and the difference is stated rather than averaged out.
| Service | May they train on what is sent? |
|---|---|
| Google Gemini | Yes, on the free tier this uses. Google's terms for it allow using API content to improve their products. A paid tier does not. |
| Ollama Cloud | No. Their policy states inputs and outputs are not used to train any model, and that cloud content is processed transiently. |
| Groq | No. Not permitted to use API inputs or outputs to train models. |
Which of them answers a given request depends on which is configured and which is available at that moment; there is no way to choose one from the page.
If that matters to you for a particular chat, the second option under the reading button runs a model on your own computer — Ollama or LM Studio — and sends nothing to anybody. It is slower and needs a one-line setup, and it is the option we would pick for a chat we cared about.
You can also press Counted only on any reading. That gives you the same numbers and the same exact lines with no model involved at all, and nothing leaves the page.
A published quiz — the one thing we store, and only if you press the button
Everything above is about reading a file. Publishing a quiz is different, and it is the only part of this product that keeps anything. If you never press Send this quiz to the group, nothing on this page applies to you and nothing about you exists on our side.
When you do press it, the dialog shows you the exact record that will be stored, in full, before it is sent. That record is:
- The questions, their four options each, which option is correct, and the host's lines.
- The group name and one-line description the model wrote for the quiz.
- The lines the quiz quotes, and only those. A Hot Seat question carries the line from your chat that proves its answer; in Guess who said it the question is a line somebody typed. Those are real messages and we will not pretend otherwise — what we promise is narrower and checkable: the only messages stored are the ones printed in the publish dialog, in full, before you press the button, and they expire with the quiz. The dialog counts them for you.
- Nothing else. No export, no dates, no counts, no participant list, and no copy of the file you dropped — that file never left your browser, so there is no copy of it here to keep or to leak. Names appear only where they are already written into a question or an option, which you can read on screen before publishing.
The endpoint does not filter your chat out — it copies the fields listed above into a new record and ignores everything else, so a future bug on our side cannot start storing your messages by accident. That is a property of the code rather than a promise about our intentions, and it is enforced by a test that publishes an entire chat export and checks none of it survives.
It deletes itself. You choose 24 hours, 7 days or 30 days, and that expiry is written onto the record in the database, which removes it when the time passes. There is no option for "forever", and no cleanup job that somebody has to remember to run. You can also delete it instantly from the browser you published it in — there is no account, so that browser holds the only key, and clearing your site data means you lose the early-delete button while the automatic expiry still applies.
Scores. Anyone who plays a published quiz types a name for the leaderboard. That name and their score are stored with the quiz, are visible to anyone holding the code, and expire with it. Nothing else about a player is recorded — no address, no device, no account.
Who can find it. Only somebody with the eight-character code. Codes are random out of about a trillion, there is no way to list or search quizzes, and the page carries a no-index tag so search engines stay out of it.
The store is Upstash Redis, hosted in Mumbai, holding exactly the record described above for exactly the time you chose.
What we never do
- No account, no sign-in, no email address collected.
- No cookies set by us, and no analytics or tracking script of any kind.
- No advertising, and nothing sold or shared with anybody.
- Nothing generated by a model is ever written onto the chat card or into the video, and nothing generated is ever left unmarked anywhere. The one place model-written words appear on an image is the roast card, which you make on purpose by pressing a button: there the roast sits in its own panel, carries a label reading "written by AI", and names the model that wrote it. Every number outside that panel was counted on your machine.
Children
This is not built for or aimed at children under 13, and we collect nothing from anybody.
If this page and the product ever disagree
Tell us and we will fix whichever is wrong. Say hi — that inbox is read by a person.
Last updated 7 August 2026.