There is one test that settles the whole thing, and you can run it yourself in about a minute. Turn on airplane mode, so every radio is off, open your phone's summarize tool, paste a long, boring email, and ask for three lines. If it answers with no spinner and no "you're offline," the work happened on the device in your hand. If it stalls, it was always going somewhere else. That is the claim these 2026 phones are selling, and for once the claim is testable.
Here is what is actually new. Apple's Foundation Models framework in iOS 26 exposes a small on-device model, the kind sized for quick jobs and not much more (roughly 3 billion parameters), running on the Neural Engine, the dedicated coprocessor in the chip, not the cloud. In practice it feels close to instant, the text landing about as fast as you can read it (Apple's own machine-learning writeup puts it near 30 tokens per second on an iPhone 15 Pro), and the developer documentation is blunt that apps can call it with no network access at all. Google's side is Gemini Nano, which runs inside a system service called AICore. The Android developer docs state flatly that "AICore does not have direct internet access." Google's support page then says the quiet part out loud: features "work even when you have no network connection or are in Airplane mode," and sensitive information "is never sent to the cloud or to Google."
So the privacy win is real, and it is worth naming precisely, because it is narrow. A prompt that never leaves the device cannot be intercepted, cannot be breached in someone else's server dump, and cannot be handed over under a subpoena the company never received, because the data never reached them. That is a genuine, structural difference from typing the same thing into a chatbot that phones home. It also holds up over the two years you will actually own the phone, because it does not lean on the company keeping any promise. So ask the only question that matters here: what does this cost you in exposure? For that one task, close to nothing.
Now the part the marketing files under the fold. On-device is not the same as private. Those are two different promises, and the phones keep letting you conflate them. Samsung is the clearest example because Samsung wrote it down. Its Personal Data Engine, the thing that powers the Now Brief dashboard, is described in the phone's own Settings, as reported by SlashGear, as analyzing "calendar events, reminders, contacts, call logs, text message history, images and videos in Gallery, location, media play history," Samsung Health data, your account profile, and app usage. All of that can be processed on the device and still be a staggering amount of reading. And Samsung's own published privacy notice states that for Now brief specifically some content "may require the processing of your data outside your device, including by third parties." So the flagship "AI phone" feature is local until, quietly, it is not.
Apple is cleaner but not absolute: larger requests fall back to what it calls Private Cloud Compute, which is a server, however well sealed. The small local model is deliberately limited, good at summarizing, rewriting, quick replies, and short back-and-forth, and by Apple's own account specialized for everyday text tasks rather than general world knowledge. The moment you ask it something big, the phone wants the network again.
Which is why the airplane-mode test is the only demo I trust, and why you should run it before you believe anyone. Turn the radios off, open the feature, give it a real prompt. If you get a real answer, the inference is genuinely on your device. If it stalls or errors, that almost certainly means it was routing to a server, though a stall can also just be a model still loading or a plain bug. For the programmers reading: this maps directly to the APIs. Apple's SystemLanguageModel and Google's ML Kit GenAI calls resolve locally; anything reaching for Private Cloud Compute or Now Brief's off-device content does not, and you can watch which is which.
The two-year question is where I land. Local inference is the rare feature that gets more valuable as the company gets less trustworthy, because it does not depend on the company staying good, or staying alive. Use it, and keep the network off when you can. Just do not let "runs on your phone" launder "reads everything on your phone." Turn on the on-device-only toggle where the phone offers one, and assume the assistant that drafts your replies is the same assistant that has already read your calendar. The next thing they sell you will be a subscription to summarize it faster.
