A journal used to be paper, and paper kept secrets by default. Now wellness apps invite you to type your inner life into a text field, and the companies behind them are — like everyone — racing to add AI. Those two facts meet in one question worth asking before your next entry: is what I write here used to train something? You can usually find the answer in a few minutes, in the privacy policy rather than the marketing. This guide shows you where to look and how to translate what you find.
Why this question, now
AI features run on models, and models get better on exactly one diet: data. Text written by real people about their real feelings is unusually nourishing data — and a journal is that, in concentrated form. So a wellness company holding millions of entries holds something its AI team can't help noticing.
Noticing an incentive is not an accusation. Plenty of companies hold that data and never touch it for training; some ask permission properly; a few write themselves broad rights and hope nobody reads clause 4.3. The point is that you can't tell which kind you're dealing with from the app's calm lavender branding. You can tell from the policy — because the broad rights, to be usable, have to be written down.
Find the AI clauses in two minutes
Open the app's privacy policy in a browser and search the page (Ctrl-F, or ⌘-F; on a phone, "Find in page"). Try each of these, in roughly this order:
- "train" — the frankest word, when it appears.
- "machine learning", "artificial intelligence", "model", "LLM"
- "improve our services" — the phrase training most often lives inside.
- "de-identified", "anonymised", "aggregated"
- "service providers", "affiliates", "partners" — who else touches the data.
- "opt" — as in opt-in or opt-out, and which one you got.
Read the full sentence around each hit, and note the policy's "last updated" date while you're there — AI clauses are usually recent arrivals, and a quietly updated policy is how they arrive. If a search for "content you provide" or "user content" leads to a licence grant, read that twice.
Translating the soft phrases
- "We may use your information to improve our services." The broadest verb in the genre. Unless the policy explicitly excludes model training, this sentence is wide enough to contain it — that width is usually the point.
- "Only in de-identified or aggregated form." For counts and averages, this means something. For free text it is weaker than it sounds: prose is made of who-you-are — names, places, the specific argument, the town's one hospital. Stripping the email address off a diary entry does not make it anonymous; researchers re-identify "anonymised" text with depressing regularity. And de-identified or not, your sentences would still be inside the model.
- "Shared with service providers and affiliates." Your entries may be processed by companies you've never heard of, under contracts you'll never see. The better policies name their AI vendors; silence about who is a kind of answer.
- "You may opt out in settings." Better than nothing — and still a tell. Opt-out means the default, chosen for everyone who never finds the toggle, is yes.
The honest caveat, as ever: a broad clause is not a confession. Legal teams write defensively, and some well-behaved apps sit under maximalist policies. But you aren't being asked to trust the team's intentions; you're being asked to agree to the document. Read the document.
Four questions that deserve plain answers
- Where does the AI actually run — on my device, or on a server?
- Does my journal text ever leave the phone, even encrypted, even "securely"?
- Is anything I write used to train or fine-tune models — theirs, or a vendor's?
- When I delete an entry, is it gone everywhere — or only from my screen?
An app that's proud of its answers puts them where you can find them. Vagueness on all four at once is its own information. For the first question, we've written a gentle explainer on what on-device AI means and how to check for it — and the two-minute privacy-label check covers the wider habit.
The architecture that ends the question
There is a design where this entire article becomes unnecessary, and it's the one we chose for Maitora. Mitra, the app's AI companion, runs on your phone — the model ships inside the app and does its thinking on the device, in airplane mode if you like. Your journal is stored on your phone. No Maitora servers exist for it to reach.
So our answer to the training question isn't a promise of restraint. It's an inability: we cannot read your journal, so we cannot train on it — nobody at Maitora could get your entries if they wanted to, and a policy update can't quietly change what the architecture makes impossible. That's what "private by design" means when it's structural rather than decorative.
Structural, not promised. No accounts, no analytics, no servers — read how Maitora is built in Privacy by design.