Is an AI journal private? Five questions to ask before you tell it anything
The pause before you type the true sentence — "where does this go?" — isn't paranoia; it's good judgment. Here are five questions to ask any AI journal before you trust it, how to answer them without a law degree, and how we answer them ourselves.
The cursor is blinking. You have the sentence ready — the real one, the one you haven't said out loud to anyone — and your hands are on the keys. And then the other thought arrives, quiet and reasonable: where does this go?
Most writing about journaling apps treats that thought as an obstacle to be soothed away. This piece treats it as what it is: good judgment. You were about to hand something intimate to a system you cannot see, run by people you have never met, governed by a document you have never read. Hesitating is not paranoia. It is the correct response to an unfamiliar room. The goal is not to make the hesitation disappear — it's to answer it properly, so that when you do write, you write freely.
Here are the five questions worth asking of any AI journal — ours included — and how to find the answers without a law degree.
1. Is your writing used to train models — and how hard is it to say no?
The most consequential sentence on any AI journal's privacy page is the one about training. If your entries feed model training, fragments of your inner life become raw material for a system that serves strangers — abstracted, statistically diluted, but yours. Some companies do this and say so honestly; some do it and hide the exit.
How to check: open the privacy page and search for "train" and "improve." The honest phrasing is a sentence: we do not use your content to train models, full stop. The evasive phrasing is a setting — an opt-out three menus deep, switched on by default, described as "helping us improve our services." A setting is not a policy. If the answer to this question is a toggle, notice who chose its default for you.
2. Is it sold or shared with data partners?
Almost no app sells "your journal entries" outright — that is not how the trade works. The trade works through softer words: , , , . "We may share information with trusted partners" is a sentence that can mean nearly anything, which is precisely why it gets written.
affiliates
partners
service providers
advertising
How to check: search for "sell," "share," "partners," and "advertising." Watch especially for "de-identified" or "aggregated" — de-identified data has a well-documented habit of becoming re-identified. And ask the plain business question, with no villainy implied: where does the money come from? A journal that is free forever has to pay its servers somehow. Sometimes the answer is patience and venture capital. Sometimes it's you.
3. Can you actually delete it — or only archive it?
Deletion and archiving look identical from your side of the screen: the entry disappears. Underneath, they are different acts. "Deleted" sometimes means removed from view while retained in backups, analytics tables, or "for legal and operational purposes" for an unstated time.
Honest retention exists — backups genuinely take time to cycle out, and a page that says deleted within 30 days from all systems is telling the truth respectfully. What you're checking for is whether deletion is something you can do, now, with a button inside the product — or something you must request, by email, to be processed eventually. Search for "retain" and "delete." The distance between those two experiences is the distance between owning your words and borrowing them back.
4. If you speak, where does your voice go?
Voice is more intimate than text. A recording doesn't just carry your words; it carries your tone at one in the morning, the pause before the hard part, the crack in the middle of it. So the technical question matters: when you speak to the app, is your speech turned into text on your own device — or is the audio uploaded to a server, and if so, is it kept?
How to check: search the policy for "voice," "audio," and "recordings." A sentence like "we may retain audio to improve speech recognition" means your voice has become training data — see question one, now in a register no one can imitate but you.
5. Whose infrastructure holds it — and can they read it?
Every journal lives on someone's computers. The question is whose, and how many hands your words pass through on the way. An app assembled quickly on rented services may route your entries through more parties than you'd guess: analytics kits, error trackers, third-party processors. None of this is sinister — it is simply how much modern software is built — but every additional hand is one more privacy policy you are trusting without having read it.
How to check: look for a "subprocessors" or "third parties" section. A company that publishes the list at all is showing you something; the length of the list is roughly the number of promises you're actually relying on.
The uncomfortable part: no app deserves blind trust
Including ours. Reassuring sentences are cheap to write, and every company writes them. What separates an answer from a vibe is checkability: narrow, factual claims stated plainly — we do X, we never do Y — rather than broad emotional ones — your privacy is our priority. The first kind can be caught lying. The second cannot even be pinned down. You will never personally audit anyone's servers, so choose the products that can be caught.
How Arkhetia answers the five questions
Since we're handing you questions to point at everyone, here is how we answer them ourselves — in the checkable register we just asked for:
Training: your words are never used to train AI. There is no toggle to find, because there is nothing to opt out of. Selling: never sold, never shared with data partners. Deletion: deletable in a click — a session you remove is removed, not archived out of sight. Voice: if you speak instead of type, your voice never leaves your device; what travels is the text, not the sound. Infrastructure: your sessions live on infrastructure we control. And one answer you didn't ask for but belongs to the same lock: the AI that reads your session is never sent your account name or your email — only a session name, and only if you've explicitly chosen to share one.
We built it this way for a reason that is more practical than noble. Arkhetia is the AI journal that doesn't just agree with you — sessions with a beginning and an end, what you bring read through three lenses, what you explore accumulating into something over time. None of that works on guarded sentences. The product only earns its keep if you write the true one — and nobody writes the true sentence into a room they don't trust. (And to say it plainly, as we always do: it is not therapy, and privacy is not the only thing to check before you lean on any tool this way — [here is our honest map of those limits](/library/en/talking-to-chatgpt-about-your-problems).)
Check the lock before you speak
Writing honestly requires a room you trust — that is not a metaphor about software, it's a fact about people, older than software. The hesitation you felt at the blinking cursor was the right instinct. Don't override it; answer it. Spend ten minutes with the privacy page of whatever app is holding your inner life, run the five searches, and see what kind of sentences come back. Then, in whichever room passes — write the sentence.
Reading about a pattern is one thing. Seeing where it runs your own life is another. Arkhetia works through these lenses — with you.