AI Tarot Apps: Why On-Device Matters More Than the Model

"AI-powered" is now on the icon of half the divination apps in the store. Almost none of them tell you where your question goes after you press send. That detail matters more than which model answers it — because the questions people ask an oracle are not the questions they'd put in a search engine.

What you actually type into an oracle

Look at your own search history and you'll find recipes, directions, and spelling checks. Now imagine the log of everything you've ever wanted to ask a fortune teller. Whether to leave. Whether they meant it. Whether the lump is nothing. Whether you'll be all right.

That's a different category of data. It isn't "personal information" in the sense of an email address — it's the shape of someone's private worry, in their own words, timestamped. And in most AI tarot apps it travels to a server, gets logged, and sits in a database whose retention policy you have never read.

The three ways an AI reading can be built

A written set of readings. The app picks from text a human wrote in advance, usually keyed to the card and sometimes to a hash of your question. Nothing is sent anywhere. Nothing is generated either — ask twice and you'll notice. This is what most apps called "AI" were doing before 2023, and a surprising number still are.

A cloud model. Your question goes over the network to OpenAI, Anthropic, Google or a reseller, and the answer comes back. The writing is genuinely responsive, which is why it feels like magic. The cost is that your question left your phone, and you're trusting at least two companies — the app developer and the model provider — with what you asked and when.

An on-device model. The model runs on the phone's own silicon. The question never touches a network; there's nothing to log, because there's no server in the loop. The trade-off is real: an on-device model is smaller than a datacentre one and writes accordingly.

Why we picked the smaller model

Enigma's Oracle uses Apple's on-device Foundation Models, built into iOS. Your question is composed into an answer on the phone and never sent anywhere, and the app has no server, no account, and no analytics at all.

The honest cost: a phone-sized model writes shorter and plainer than a cloud one. It won't produce three paragraphs of ornate prose. Given the choice between a more eloquent answer that leaves the device and a plainer one that doesn't, the second is the right trade for this kind of app. For a reflection prompt, "plainer" is arguably better anyway — the value was never in the eloquence.

How to tell which kind you're using

• Turn on airplane mode and ask. The fastest test there is. If an answer still appears, nothing was sent. If you get an error or a spinner, it was going somewhere.
• Read the App Store privacy card. Scroll to "App Privacy" on the product page. An app that ships your questions to a model provider has to declare it — and if it declares nothing while advertising AI, that's its own kind of answer.
• Ask the same question twice. Word-for-word identical replies mean written readings, not generation, whatever the icon says.
• Look for an account requirement. Sign-in usually means a server, and a server means your questions have somewhere to live.

Where on-device genuinely falls short

It would be dishonest to pretend the trade-off is free. On-device models need recent hardware, and Apple Intelligence has to be switched on. It also doesn't speak every language yet — Ukrainian and Russian among them — so in those languages Enigma falls back to its written readings and says so on screen rather than pretending the model answered.

A cloud app has none of those limits. It works on an old phone, in any language, with the same fluency everywhere. If that matters more to you than where the question goes, that's a legitimate choice — just make it knowingly, which is the whole point of this post.

The part that isn't about AI at all

None of this makes a reading more accurate. A card is a prompt for reflection, and the model — local or remote — is a mirror with reasonable lighting. What on-device generation changes isn't the quality of the insight; it's whether asking the question costs you anything.

If you're new to reading cards at all, start with the ritual rather than the technology: our guide to the daily one-card draw needs no app, and the psychology behind why it works has nothing to do with models either.