Is it safe to tell an AI your dreams?

It depends entirely on where your dream goes after you tell it. A dream spoken or typed into an app may stay on your phone, sit on a company’s servers indefinitely, or train someone’s model — and the marketing rarely says which. Three questions, answerable in about two minutes, tell you.

That’s the short answer. The longer one is more interesting, because “is it safe” turns out to be a question about plumbing — and the plumbing, unlike the marketing, is checkable.

What actually happens when you tell an app your dream?

More than you’d guess, and most of it out of sight. Between you telling a dream and an interpretation appearing on screen, the dream makes a small journey: it is captured, transcribed if it was spoken, usually transmitted, possibly stored, and — under some terms of service — retained for purposes that have nothing to do with you, including training future models. Every step happens in one of two places: on your phone, or on someone’s server. “Is it safe” reduces to knowing which, at each step.

The app’s homepage will not tell you. Homepages say “completely private” the way restaurants say “world famous”. The honest answers live in two dull documents — the privacy label on the app’s store page and the privacy policy behind it — and almost nobody reads either. In a Pew Research Center survey published in October 2023, 56% of Americans said they frequently agree to privacy policies without reading them, and 67% said they understand little to nothing about what companies do with the data they collect — up from 59% in 2019. We are disclosing more and understanding less about where it goes.

Most of what travels through that gap is mundane — locations, purchases, step counts. A dream is not mundane, which is why it deserves a closer look before it makes the trip.

Why do dreams deserve more caution than other data?

Because a dream has always had a chosen audience. In surveys reviewed by Blagrove and colleagues in a 2019 Frontiers in Psychology study, about a third of people share a dream at least monthly and roughly one in ten do so weekly — yet only around 14.5% of dreams are ever shared, almost always with a romantic partner, a close friend, a relative. Read that from the other side: some 85% of remembered dreams are never told to anyone. You have been curating this audience your whole life, mostly without noticing — the odd little dream for the breakfast table, the stranger one saved for the friend who won’t laugh, the third kept entirely to yourself.

An app asks to join that inner circle, and it asks at your least guarded hour. Dreams have to be caught in the first minute of waking or they dissolve, so you tell them before you’re awake enough to be careful. And unlike the friend at the kitchen table, a company can’t be read across the table. You know what your friend will do with your dream because you know your friend; what a company does with it is decided in documents, and whether it keeps its word is invisible from the outside. People sense this: in the same Pew survey, 70% of those familiar with AI said they have little to no trust in companies to make responsible decisions about how they use it. The habit of telling dreams is older than writing. The habit of reading data policies is not going nearly as well.

How do you check what a dream app really does?

You don’t need to trust anyone, and you don’t need to read forty pages. The three questions from the top — where is the dream transcribed, where is it stored, does it train anything — are answered by five checks:

1. Open the privacy label before you download. Since December 2020, Apple has required every new app and update to declare what data it collects and how it is used; the result sits on every App Store product page under “App Privacy”. It is self-reported, but developers are responsible for its accuracy, which makes it the closest thing to a sworn statement you will get from an app.
2. Read the label against the table below. The four headings mean specific things, and for a journal the differences between them are enormous.
3. Search the privacy policy for three words. Open the policy and search for “train”, “retain”, and “third”. You are not reading the policy; you are auditing it. A provider that trains on user content, keeps it indefinitely, or shares it with third parties has to say so somewhere, and those three words find the sentence faster than any amount of scrolling.
4. Ask where transcription happens. If you speak your dreams, this is the whole game. On-device transcription means the text is produced on your phone and the audio never needs to travel. Server-side transcription means a recording of your voice, saying your dream, exists on someone else’s machine — however briefly, on their terms. An app that won’t answer this question plainly has answered it.
5. Check whether it demands an account. A journal tied to your email address is a journal tied to you, everywhere that data flows afterward. No account means there is no “you” in the database for dreams to attach to.

Label	What it means for a dream journal
Data Used to Track You	Something from the app is linked with third-party data for advertising. In a journal, disqualifying.
Data Linked to You	Entries or usage are stored connected to your identity — account, device, or other identifiers. Your dreams sit in a database that knows your name.
Data Not Linked to You	Data is collected but stripped of identifiers first. Better — though “not linked” is a promise about processing, not a limit on collection.
Data Not Collected	The developer takes nothing from the app. The only category that is an architecture rather than a promise.

What would a genuinely private dream app look like?

Not like a promise; like a floor plan. The criteria are concrete, and you can hold any app to them:

• Transcription happens on the phone. The dream becomes text on the device itself, so no recording of your voice ever needs to leave it — and the audio file can be deleted the moment the text exists.
• The journal lives in your cloud, not the company’s. On iPhone, that means the user’s private iCloud database, readable only by their own Apple account — the developer cannot open it.
• The server keeps nothing dream-shaped. If interpretation involves a server at all, the dream text is sent for that single request, processed in memory, returned, and gone — not stored, and never written to a log.
• There is no account. Nothing to sign up for, nothing to leak, no email address for a database of dreams to orbit.

This is the standard Dreamsay was built to — it is the reason the app exists in the form it does, and the reason its central promise can be checked rather than believed: your voice never leaves your device.

One quiet consequence of that kind of architecture is worth noticing. A company that never receives dreams cannot aggregate them, sell them, lose them, or publish a report on what ten thousand of them contain. Whatever patterns your journal holds belong to the only database that has them — yours.

Whose ears were they for?

The question “is it safe to tell an AI your dreams” sounds new, but you have been answering its older form since childhood. Every morning retelling was a decision: this dream for the breakfast table, that one for a particular friend, the rest for no one at all. Nobody taught you this policy. You wrote it yourself, dream by dream, and it has served you for decades.

So keep writing it. A dream journal is worth keeping — your own dated record will tell you more than any dictionary or any model — and an app earns a place in your mornings the way a person earns a place at your table: by showing you, not telling you, what it does with what it hears. Two minutes with a privacy label is a small price for knowing exactly whose ears your dreams are for. Yours first. Everyone else is your call.