Before trusting an AI journal with sensitive memories, get answers to four things: where entries are stored, whether they train a model, who inside the company can read them, and what deletion actually removes. Then check two more that matter just as much, what the product is designed to optimize for, and how it behaves when you write something difficult. Most of this sits in the settings screen and one page of the privacy policy.
Ask about the data and the AI
Find out where entries are stored, whether they are encrypted, which AI providers process them, and whether content is used for training. Ask how long deleted data remains in backups and whether you can export your journal in a usable format.
Encryption is worth a follow-up question rather than a checkbox. Encrypted in transit and at rest is normal and good; end-to-end encrypted means the provider cannot read your entries, which is a stronger and much rarer claim, and one that is difficult to combine with AI features.
- Can employees or contractors read entries?
- Can I turn memory features off or correct them?
- What happens to my data if I close the account?
- How does the app handle an inaccurate or unsafe AI response?
- Which company actually runs the model that reads my entries?
- Can I use the journal without AI features at all?
What is the product designed to optimize for?
Look at the business model and the interface. Does the app help you finish a reflection, or does it keep manufacturing reasons to stay? Does it make clear that AI is not medical care? Can you write without accepting every prompt?
Incentives are easier to read than intentions. A journal you pay for is being paid to be useful; a free journal with no visible revenue is being paid by something you have not identified yet, and that is worth identifying before you write anything difficult in it.
You do not need perfection from a journal. You do need enough transparency to make an informed choice about the most personal material you plan to share.
How does it behave in a difficult moment?
Write one genuinely hard entry, or something that resembles one, and read the response carefully. This tells you more than any policy page.
Good behavior looks unglamorous: it stays with what you said, does not diagnose you, does not perform concern, and points toward human help when something is beyond a journal. A tool that offers confident interpretations of your mental state after four sentences has told you what it will do with four years.
Decide your own deal-breakers first
It is easier to evaluate a product when you know in advance which answer would end the conversation. For most people it is one or two things: training on entry text, or an inability to export.
Everything else becomes a tradeoff you can weigh calmly rather than a source of general unease. Vague answers to specific questions count against a product, but a clear no you can live with counts for it.
The two questions that predict the rest
Ten questions is a good checklist and a lot to hold. If you only ask two, ask these, because the answers correlate with almost everything else you cannot inspect.
First: can I export everything, right now, in a format I can open without this app? Export costs engineering time and makes leaving easier, so building it well is a deliberate choice about the relationship a company wants with you. Products that have done it tend to have made similar choices elsewhere.
Second: how does the company make money? A subscription aligns their incentives with keeping your material private and useful. Advertising or data sharing does not. This is not moralizing. It is the single strongest predictor in the category. The 2026 research finding undisclosed trackers in every one of twenty-five mental health apps examined was overwhelmingly a finding about free, ad-supported products.
If both answers are good, the remaining eight questions are usually good too.
How to test the answers rather than take them
Companies answer questions in the way that is most favorable and still true. The way past that is to check behavior instead, which takes about twenty minutes.
Run the export before you write anything meaningful and open the file. Find the delete-account control and confirm it exists as a button rather than an email address. Turn on airplane mode and see which features still work, that tells you far more about what is local than any marketing page. Search the privacy policy for "train", "subprocessor" and "retention" rather than reading it. Then write three entries and see whether anything comes back that could only have come from them.
Whatever survives those five checks is a product you can reasonably commit years of writing to. Whatever does not has told you something the FAQ was never going to.
- Export it and open the file, on day one
- Confirm account deletion is self-service
- Airplane-mode test to see what is genuinely local
- Search the policy for train, subprocessor, retention
- Three entries, then check whether the memory is real
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.