An entry typically travels further than it feels: from your device to the app's servers, often on to a model provider for processing, and back as a response: with the entry, any summary generated from it, and its embeddings stored along the way. Knowing that path turns a vague privacy worry into four specific things you can check in the settings.
An entry can be stored and processed in different places
The journal service may save your entry in its database, send some or all of it to an AI provider for processing, and keep technical logs needed to run the product. Those systems may have different retention periods and access rules.
A clear privacy notice should distinguish journal content from account details and usage analytics. It should also explain whether AI providers retain prompts, whether data crosses regions, and which subprocessors are involved.
What path does an entry usually take?
The specifics vary by product, but the shape is fairly consistent. Knowing the steps is what lets you ask a precise question instead of a general one.
- Your device holds a draft, often including autosaves and keyboard caches.
- The text travels encrypted to the journal service and is written to its database.
- For AI features, some of it is sent to a model provider, which may log it briefly for abuse monitoring.
- Backups and operational logs keep copies on their own schedules.
- Search indexes and generated summaries may exist as separate records of the same content.
Summaries and memory are data too
This is the part most people miss. If a product generates a monthly summary or maintains a memory of what matters to you, that derived text is its own stored record, written in a compressed and unusually revealing form.
It is worth asking whether deleting an entry also removes what was derived from it. Frequently it does not, which means the conclusion can outlive the sentence that produced it.
What should deleting an entry actually do?
Removing an entry from your screen does not always mean every copy disappears immediately. Backups, safety logs, and legal retention requirements can affect timing. Look for a concrete explanation of what is deleted, from where, and within what period.
A good answer sounds specific and slightly unglamorous: removed from the live database immediately, purged from backups within thirty days, excluded from memory at once. A vague reassurance that data is deleted when you delete it is not an answer to any of those questions.
If the answer is difficult to find, ask support before adding highly personal material. You deserve to know the lifecycle of words that matter to you.
The third party most people never think about
Very few journaling apps run their own models. Almost all of them send your text to a large provider, OpenAI, Anthropic, Google, and return the response. That means your entry usually touches at least two companies: the app you chose, and a model provider you did not.
The 2026 study of twenty-five mental health apps found that forty-eight percent referenced third-party AI providers in their policies. Some named them. Others used language vague enough to cover anything, like "AI services" or "third-party processors". Several disclosed using personal conversations for training or product improvement, including the possibility of commercializing datasets derived from them.
The practical consequence is that the app's privacy promises are capped by its provider's terms. An app can be entirely honest about its own conduct and still be routing your entries through infrastructure with different retention rules. When a policy names its subprocessors, that is a good sign. When it will not, you are being asked to trust a chain you cannot see.
Does deleting an entry actually delete it?
Usually, eventually, with caveats worth knowing.
Deleting from the interface generally removes the entry from your view immediately and from the primary database shortly after. Backups are the first complication: most services keep rolling backups for somewhere between thirty and ninety days, so a deleted entry commonly persists in a restorable snapshot for weeks after it disappears from your account.
The second complication is derived data. If the app generated a summary, extracted the people you mentioned, or built any kind of memory from the entry, deleting the original may not delete what was learned from it. Ask specifically whether deletion propagates to summaries and memory. Many products have not thought about it.
The third is legal hold, which overrides everything above. If a provider is under a preservation order, deletion stops for the duration, as ChatGPT users discovered in 2026 when conversations they believed erased were retained under exactly that mechanism.
How to find out for your own app
You can establish most of this in about fifteen minutes without any technical knowledge.
Search the privacy policy for the words "subprocessor", "third party", "retention", "backup", and "train". Those five terms surface the substance and skip the boilerplate. Then look at what the app actually lets you do: is there a delete-account button that works without emailing support, and does exporting produce a real file?
The gap between what a policy promises and what the interface permits is the most reliable signal in the entire category. A company that has built working export and deletion controls has spent engineering time on your exit, and companies do not do that accidentally.
Sources
- Mental health apps are collecting more than emotional conversationsHelp Net Security
This piece is part of ai journal privacy. A journal only works if you trust it. These reads cover what happens to your entries, the difference between on-device and cloud processing, and the questions worth asking any app.