Encrypted at Rest vs. We Can't Read It: The Two Claims Every Journal App Blurs
"Encrypted at Rest" vs "We Can't Read It": The Two Claims Every Journal App Blurs
Bottom line: "encrypted" and "we can't read it" are two different products, and the whole category uses one word for both. Once you can tell them apart, picking a journal app takes about ninety seconds.
I build one of these apps. Read this as a biased source with a checklist attached — the checklist works on mine too, and I'll run it on mine at the end.
The tell that started this post
Here's a live example, and I want to be careful with it, because the company involved is the one behaving well.
Mindsera is a well-regarded AI journal. Third-party reviews this year describe it as having "privacy-first architecture with end-to-end encryption." Mindsera's own help documentation says something different: your writing is encrypted at rest with AES-256 and in transit with TLS, but end-to-end encryption is impossible to use with AI right now, because AI systems need access to the data in the backend to work.
Both of those sentences are on the internet today. The engineering disclosure is the honest one. The market's summary of it is wrong.
I'm not picking on Mindsera — I'm pointing at them because they wrote down the true thing. Most apps in this category don't write it down at all. They say "encrypted," let you fill in the rest, and never correct the reviews.
Three tiers, not two
The word "encrypted" is doing the work of three very different promises.
Tier 1 — encrypted at rest. Your entries are stored scrambled on a server. The company holds the key. This protects you from someone stealing a hard drive out of a datacenter. It does not protect you from the company, from an employee with database access, or from a subpoena served on the company. The lock is real. They kept a copy of the key.
Most journal apps live here. Most of them say "encrypted" and stop.
Tier 2 — end-to-end / zero-knowledge. Your entries are encrypted on your device with a key only you hold. The server sees ciphertext and nothing else. Nobody at the company can read your journal — not for troubleshooting, not under a court order, not if they wanted to.
This is the strongest promise available, and it has a hard cost: a server that can't read your entries can't analyze them. No server-side AI. No cloud search. That's not a policy choice, it's arithmetic. This is exactly what Mindsera's help page is telling you.
Tier 3 — the one nobody has a name for. The server can technically read your entries, because that's how the AI works — but the company has built and enforced a specific, testable barrier so that no human at the company can. Not a promise. A mechanism.
Tier 3 is where every honest AI journal actually sits, whether or not it admits it. The interesting question isn't whether an AI journal touches your plaintext somewhere. It does. The question is what stands between your plaintext and a person.
Why "on-device AI" isn't a free pass either
A handful of apps now run analysis on your phone and pair it with zero-knowledge storage. That's a real architecture and I respect it. It's the honest Tier 2 answer, and if it's what you want, take it.
The tradeoff is capability. A model small enough to run on a phone is not the model that writes you a useful letter about your month. You're choosing a weaker reflection in exchange for a stronger guarantee. That's a legitimate trade — it's just a trade, and the marketing rarely says so.
Anyone telling you they've got frontier-quality AI reflection and true zero-knowledge storage, with no tradeoff at all, is describing something that doesn't exist yet.
The four questions
You don't need to understand cryptography. You need four questions that force an app off the word "encrypted."
1. Who holds the key? If the answer is "we do, securely" — that's Tier 1. If it's "only you, derived on your device" — that's Tier 2. If they won't say, assume Tier 1.
2. Could an employee read my entry if they tried? Not "would they." Could they. Push until you get a yes or a no. "We take privacy seriously" is a non-answer and you should treat it as a no-comment.
3. If the answer is no — what makes it no? This is the question that separates a claim from a guarantee. Is it a policy in a handbook? An access-control setting? Or something that fails a build, breaks a test, and stops a release? Policies change quietly. Tests don't.
4. What happens to my words after the AI reads them? Discarded, or retained to improve the product? "You can opt out" means the default is that they use it. "Never" is a different sentence. Ask which one is in the contract with their AI providers, not just the marketing page.
That's it. Four questions, any app, ninety seconds.
Running it on Uncloud
My turn. Uncloud is a private AI journal that turns what you write into insight — offline-first, yours alone. Here's where it honestly sits.
Uncloud is Tier 3. It is not end-to-end encrypted and it is not zero-knowledge. I'd rather say that in my own blog post than have you find it out later. The AI reflects on your text, which means your text is processed. Any app that reflects on your writing with a serious model does this, and the ones that imply otherwise are selling you a word.
So, the four questions:
Who holds the key? Entries are written offline-first on your device, and the canonical copy lives there. Sync is envelope-encrypted in transit and at rest, with per-user keys held in a separate key-management system.
Could an employee read your entry? No. Our admin tools are blocked, in code, from reading entries.
What makes it no? That block is enforced by automated tests in every build. If a change ever broke it, the build fails and the release stops. It isn't a policy I'm asking you to trust; it's a test that has to pass before the app ships. Our Play listing puts it as: encrypted so completely that even we can't read it, enforced by automated tests in every build.
What happens to your words? Your journal is never used to train AI models. That's contractual with our processors — not a setting you have to find and switch off.
And where Uncloud loses points, since a checklist you always pass is a marketing brochure: we don't have a third-party security audit badge. The guarantee lives in the architecture and the test suite, not in an auditor's PDF. If an independent audit is your non-negotiable, that's a real reason to choose otherwise, and I'd rather say it plainly.
The part that actually matters
There's a reason this vocabulary problem persists: the honest answer is longer than the dishonest one. "Encrypted" fits in a store listing. "Encrypted at rest, keys held by us, readable by staff for support" does not.
But your journal is the most honest thing you write. It's the place you put the sentence you wouldn't say out loud. That deserves better than a word that's been sanded down until it means "we tried."
So ask who holds the key. Ask whether a person could read it. Ask what makes the answer no. Then pick whichever tier you can actually live with — including Tier 2 with a smaller model, if that's your line. Any of those is better than assuming.
The worst outcome isn't choosing Tier 1. It's thinking you chose Tier 2 because a review said so.
Uncloud is live on Google Play, free to start: https://play.google.com/store/apps/details?id=you.uncloud.app
iOS is in App Store review.
Uncloud is a journaling and self-reflection tool. It is not a medical device and not a substitute for professional medical, psychological, or mental-health care.