Guide
Does Apple Intelligence Work Offline? The Airplane-Mode Test
Ask Siri something in a subway tunnel and sometimes you get an answer, sometimes you get an apology about needing an internet connection. Both are correct behavior. Apple Intelligence was never designed to be all-local or all-cloud — it is a hybrid system, and Apple says so openly. What Apple does not publish is a complete feature-by-feature routing table. So the honest way to find out what works offline on your device, running your OS version, is to test it. That takes about two minutes and no technical skill. This guide covers what Apple documents, how to run the test, and how to read the results.
The short answer: partly — and Apple documents which parts
Apple's stated design is that Apple Intelligence handles many requests entirely on-device, using foundation models that run on the Neural Engine of Apple Intelligence-capable hardware. Requests that need more computing power than the device can supply are sent to Private Cloud Compute (PCC) — Apple-operated servers built on Apple silicon, designed so that your data is used only to fulfill the request and is not stored afterward. A third route exists only if you turn it on: the ChatGPT extension, which sends what you explicitly share to OpenAI.
So "does Apple Intelligence work offline?" resolves to: the on-device portion works offline; the PCC portion and the ChatGPT portion do not, by design. The interesting question is which features live on which side of that line — and because Apple can rebalance the split with any OS update, the airplane-mode test below is more reliable than any static list, including the one on this page.
On-device, Private Cloud Compute, or ChatGPT: the three routes
- On-device. The model runs on your iPhone, iPad, or Mac. Nothing leaves the device, which is why this route keeps working in airplane mode. It is also the only route whose privacy you can verify yourself with no trust in anyone — if the radio is off, the data cannot travel.
- Private Cloud Compute. A network request to Apple's servers. Apple's published commitments are unusually strong for cloud AI: stateless processing, no data retention after the response, no privileged Apple access, and server software images made available for independent security research. It is still a server, though, so it fails cleanly when you are offline.
- ChatGPT. Off by default, opt-in, and by default asks before each request it hands over. Anything sent goes to OpenAI's servers under OpenAI's policies, not Apple's. Obviously requires a connection.
One thing worth stating plainly, because this site's whole premise is verification: the operating system does not show you a per-request badge saying "this one ran locally." Airplane mode is the user-accessible instrument. Whatever still works with the radios off is provably on-device; whatever errors out needed a server. That test requires zero trust in Apple, in app developers, or in this page.
The two-minute airplane-mode test
- Go offline for real. On iPhone or iPad, open Control Center and enable Airplane Mode — then check that the Wi-Fi icon is actually off, because iOS remembers if you previously re-enabled Wi-Fi during airplane mode. On a Mac, turn off Wi-Fi from the menu bar and unplug Ethernet.
- Confirm you are offline. Open any webpage; it should fail. Now every AI feature that still works is running on your hardware, full stop.
- Try the feature you care about. Select text and open Writing Tools. Ask Siri to set a timer, then ask it a general-knowledge question. Open a long email thread and tap Summarize. Note what succeeds and what produces an error or an endless spinner.
- Turn the radios back on and retry anything that failed, so you can tell "needs a server" apart from "was broken anyway."
Two caveats that keep the results honest:
- A one-time download is not the same as cloud processing. Some features (Image Playground is the clearest example) download their models the first time you use them. If a feature has never been run on your device, an offline failure may mean "assets not downloaded yet," not "runs in the cloud." Use the feature once while online, then retest.
- Results have a shelf life. Apple moves features across the on-device/PCC boundary as its local models improve. A feature that needed a server last year may be local now, and new features may launch cloud-first. Rerun the test after each major OS update.
Feature by feature: what each one needs
Here is how the commonly used features behave on current releases (iOS/iPadOS/macOS 26). Treat this table as a starting point and the airplane-mode test as the authority.
| Feature | In airplane mode | Why |
|---|---|---|
| Writing Tools: Proofread, Rewrite, tone changes | Works | Handled by the on-device foundation model. |
| Writing Tools: "Describe your change" and heavier rewrites | Sometimes fails | More complex requests can route to Private Cloud Compute. |
| Notification and Messages summaries, priority notifications | Works | Runs on-device; these appear even with no connection. |
| Mail and Safari summaries of long content | Varies | Short content typically summarizes locally; long or complex material may need PCC. Test with your own mail. |
| Siri: timers, alarms, settings, launching apps | Works | On-device speech recognition and device actions; much of this predates Apple Intelligence. |
| Siri: general-knowledge and web questions | Fails | Needs Apple's servers, web search, or (if enabled and confirmed) ChatGPT. |
| Image Playground, Genmoji, Photos Clean Up | Works after first use | Generation runs on-device; the models download once, which needs a connection. |
| Visual intelligence lookups (search, product info, ChatGPT) | Fails | The lookup providers are online services by nature. |
If your specific question is about documents — "can it summarize this PDF?" — that case has enough wrinkles that we gave it its own guide.
Why some features can't run locally — and why that's not automatically bad
The constraint is physics, not stinginess. A phone has a few gigabytes of memory shared with everything else you're running; a server rack does not. Apple's on-device foundation model is deliberately small so it can respond instantly without flattening your battery, and small models genuinely cannot do everything large ones can. Open-ended world knowledge, very long inputs, and heavier generation are exactly the workloads that overflow a local model's capacity. (For a grounded look at what small local models are good at, see our guide to on-device AI — the short version is that they excel when the answer is in material you already have, and struggle when it has to come from the model's own memory.)
And the cloud half of the hybrid is not automatically a privacy failure. Private Cloud Compute is, by published design, unusually privacy-conscious for server AI: stateless, non-retaining, and open to outside inspection. If a feature must use a server, this is a good server for it to use.
But there is a real difference in kind, not just degree, between "verified by design documents" and "verified by your own hands." PCC's guarantees are Apple's engineering promises, checked by researchers you have to trust. Airplane mode is checked by you, in two minutes, every time. For casual use the distinction rarely matters. For a confidential contract, a medical record, or a client file, it can matter a great deal — which brings us to the workflow question.
A document workflow that is fully offline, end to end
If your motivation is privacy rather than curiosity, "partly offline" is not a spec you can build on — you want every step on the local side of the line, and you want to check it rather than take a marketing page's word for it.
That is the design goal of OpenIntelligence, the document-AI app built by this site's author. You import PDFs, Office and iWork files, spreadsheets, scans, notes, even audio and video recordings; it indexes everything locally; and when you ask a question, retrieval, verification, and answering all run on the device. Answers come with tappable citations back to the exact source passage, and when your documents don't support an answer, it says so instead of guessing. Because every route runs on-device today, the entire app passes the same test this page just taught you — and you should actually run it:
- Put the device in airplane mode (Wi-Fi off too).
- Import a document and let it index.
- Ask a question. You get an answer with citations — or an honest abstention — with the radios off.
Two checkable facts before you bother downloading. First, the requirements: it needs iOS, iPadOS, or macOS 26 or later on Apple Intelligence-capable hardware — an iPhone 15 Pro or later, or an iPad or Mac with an M1 chip or later — because it is built on Apple's on-device foundation models. Second, its App Store privacy label reads "Data Not Collected," a declaration you can verify yourself in about thirty seconds on the store listing; this guide explains exactly what that label does and does not promise. The app is free to download on the App Store, with a one-time Lifetime unlock or a Pro subscription for the full feature set.
What changes with iOS 27 — and what won't
Apple ships its major OS releases each September, and two things are worth expecting from the next cycle.
First, for Apple's own features, the on-device/PCC boundary will keep moving. Every generation of the local models takes on work that previously needed a server, and every new headline feature may launch needing one. Nothing about the airplane-mode test goes stale — it is precisely the tool for re-checking after an update.
Second, for OpenIntelligence specifically: optional Private Cloud Compute support is built and planned to go live when iOS and macOS 27 arrive, expected in September 2026. Today, every route runs on-device — nothing uses PCC yet. When the option ships, it will be off by default, consent-gated per request, used only for the final answer-writing step — never for retrieval, indexing, or your document library — and it will show you the exact payload before anything leaves the device, so the same verify-it-yourself standard applies.
What won't change, in either case: whatever runs with the radios off is yours alone. That is the one privacy guarantee no update, no policy revision, and no server operator can quietly amend — and it is two minutes away from being checked, any time you like.