On-device AI vs. cloud AI for flashcards
Every study app on the App Store right now is racing to bolt AI onto flashcards. Quizlet has Q-Chat. RemNote has AI generation. Mochi has AI grading. AnkiPro has AI card creation. Knowt, Brainscape, Studyfetch, Wisedude — pick a name and there's an "AI Powered" badge somewhere on their listing.
Almost all of them have the same architecture. You paste notes or a PDF. The app uploads it to a server. The server forwards it to OpenAI or Anthropic. The model generates cards. The cards come back. You pay a subscription to cover the recurring API cost.
Discito's AI does the same thing — generates cards from your notes, picks smart multiple-choice distractors, even turns lecture audio into a deck. The architecture is different. Your notes never leave your iPhone. Here's why we made that bet.
What gets uploaded when "cloud AI" reads your study notes
Be honest about what's in a typical study deck. A med student preparing for boards has cards on patient presentations, drug mechanisms, differential diagnoses, mnemonic patterns built from years of clinical reading. A law student has annotated case summaries, exam-style outlines, the specific framings their professors prefer. A language learner has personal vocabulary tied to specific cultural contexts, mistakes they keep making, embarrassing phrases they don't want forgotten.
That content is unusually personal in ways that aren't obvious until you imagine it leaving your device. Cloud-based AI study apps are necessarily collecting study notes from every paying customer — for the simple reason that an HTTPS upload to api.openai.com is how the model gets the text in the first place. Some apps store the content. Some apps don't. Some apps say they don't and might be telling the truth. Some apps will be acquired in three years by someone with a different policy. None of those things are knowable from the outside.
The thing about uploading your med-school notes to a third-party AI provider is that you cannot un-upload them.
The cost structure tells you why this matters
If you're a study app paying OpenAI by the token, every AI-generated card costs you money. A 50-card deck generated from a PDF might cost ten or twenty cents in API fees. Multiply by every active user, every active month, every time someone uses the AI feature, and you have a real recurring expense.
There are only three ways to cover that cost. Charge a subscription. Limit the AI feature to a few generations per day. Sell the data. Most cloud-AI flashcard apps pick combinations of all three: monthly subscription, daily AI quota, and quiet data-collection clauses in the privacy policy. None of those are villainous decisions — they're rational responses to a per-request marginal cost. But they're not the experience you actually want.
What you want is: paste your notes, get cards, do this whenever, don't think about it, don't get billed every month.
iOS 26 and Apple Intelligence quietly changed this
In 2024, Apple shipped Foundation Models — a small but capable language model that runs entirely on the Neural Engine of recent iPhones and iPads. Not "edge compute" with a fallback to the cloud. Not "private cloud" where Apple still sees the data. Actually on the device, executed locally, no network call required, no API key needed, no per-token charge.
Apple also shipped ImageCreator — on-device image generation. And in iOS 26, SpeechAnalyzer — an on-device automatic speech recognition model that handles long-form audio (hours of lecture, not just dictation snippets) with quality competitive with the major cloud transcription services.
Each of these is the building block of an AI feature Discito ships:
- AI card generation from PDFs and pasted text uses Foundation Models. You give it source material, the model drafts question-and-answer pairs in your language, you review and accept what's useful.
- AI image generation for cards uses
ImageCreator. The 4-up streaming grid surfaces variants directly in the card editor. You pick what fits. - Smart MCQ distractors use Foundation Models with a quality gate (Levenshtein-distance filter, length-band check, semantic-similarity check) to make sure the wrong answers aren't dead giveaways or rephrased copies of the right answer.
- Lecture audio → flashcards chains
SpeechAnalyzer(audio in, transcript out) into Foundation Models (transcript in, cards out). The full pipeline runs locally on a 45-minute lecture without ever opening a socket.
None of these features make a network call to Discito. None of them make a network call to OpenAI. None of them make a network call to anyone. The models live on your device; the inference happens on your device; the cards land in your Core Data store, which then syncs through your iCloud — also Apple-only, also end-to-end encrypted for most data, also not visible to Discito.
The honest tradeoff: hardware
Apple Intelligence requires specific hardware. As of writing, that's iPhone 15 Pro and later (the iPhone 17 Pro is the current ceiling), and M1 iPad and later. Older iPhones — including the entire iPhone 14 line and the non-Pro iPhone 15 — can run Discito but cannot run the AI features.
Discito handles this gracefully. The AI surfaces all show an AILanguageUnavailableBanner on unsupported devices, with a clear explanation of what's needed. Every non-AI feature works on every iOS 18.6+ device without exception — FSRS-6 scheduling, review sessions, .apkg import, iCloud sync, widgets, Live Activity, reading themes, audio playback, LaTeX rendering, the lot. Pro entitlement gates features on entitlement, not on hardware capability, except for the four Apple-Intelligence-dependent features where the gate is unavoidable.
That's not the marketing answer ("Discito requires the absolute latest iPhone"). It's the honest answer: the AI is great if you have the device for it, and the non-AI app is great on every iPhone Apple still supports.
What we don't ship and don't plan to
We don't ship a Discito cloud AI fallback. If you have an iPhone 14, you don't get AI card generation — you don't get a degraded server-side version of it either. We considered it. We decided it would be a slow leak of the whole privacy posture. The moment Discito has a server endpoint that accepts your notes, the architecture has changed, and every promise on this page becomes harder to keep.
We don't ship a "Discito web" companion that re-implements the AI features in a browser. Same reason. The browser would have to send your notes somewhere to do the inference. That somewhere wouldn't be your iPhone.
We don't have an OpenAI partnership. We don't have an Anthropic key. We don't have a relationship with any third-party AI provider, because Discito's AI features don't require one. The whole thing runs locally, and we'd like to keep it that way for as long as Apple keeps shipping models powerful enough to make it work.
The summary, in one line
Cloud-AI flashcard apps ship features that are powerful, subscription-funded, and dependent on you trusting a third party with the contents of your study deck. Discito's bet is that you'd rather have the same features without the subscription and without the third party — and that the hardware finally exists to make that the default rather than the compromise.
Privacy isn't a feature. It's a default.