An AI feature can work without an internet connection while the application around it still sends data elsewhere. “On-device” identifies where a particular computation runs. It does not describe every permission, log, backup or later action in the product.
That distinction is useful when reading the promises on an AI phone or laptop. Instead of comparing devices by a single AI label, take one feature and map its inputs, local processing, possible cloud handoff and saved output. The following worksheet is a way to ask better questions, not a product recommendation or a measured comparison of current devices.
Start with one task you can describe
Imagine a fictional notes application that turns a meeting note into three action items. The input is a note already stored on the phone. The model produces a short list. The app then saves that list beside the original.
Write those steps before asking whether the feature is private or fast. “AI assistant” is too broad to evaluate. “Extract three actions from this downloaded note” is narrow enough to investigate. Adding a web search, calendar invitation or shared workspace would create further steps with different data flows.
In this exercise, the notes application is invented. No claim is being made about how an actual vendor handles a meeting note. The point is to avoid treating several operations as one invisible event just because they share an interface.
Locate a documented local path
Google’s Gemini Nano documentation describes on-device execution through Android’s AICore service. Its listed use cases include summarization, rewriting and proofreading. It says local prompt execution avoids a server round trip, while inference speed still depends on hardware.
That supports a specific claim: some generative tasks have a documented local execution path. It does not prove that every Android device supports every feature, or that the fictional notes application uses that path. You still need the application’s documentation and the device’s compatibility information.
For the worksheet, record the evidence beside the claim. “Vendor documents a local model” belongs in one row. “This app uses it for this action on this device” belongs in another. Leaving the second row unresolved is more accurate than filling it with the first row’s answer.
Separate inference from getting ready
An offline feature may need software or model downloads before its first use. Google’s description of AICore includes managed model updates and internet requests routed through a companion service. Local execution and the distribution of the model are different activities.
For the invented notes task, ask what must already be present: the note, the required model and any application resources. If the note is stored only in an online account, retrieving it is another dependency even if the model could process it locally afterward.
A simple offline check can therefore answer only a narrow question. If the configured feature handles a downloaded note without a connection, that supports availability in that tested state. It does not establish how setup worked, what happens with a different document or whether data is transmitted after connectivity returns.
Check the handoff rather than guessing
The Android documentation also describes hybrid inference and a possible cloud fallback when the local model is unavailable or unsupported. A fallback can be useful, but it changes the question from “where does this model run?” to “which route did this request take?”
For the fictional application, add a row asking what happens when a note exceeds the local feature’s limits. Possible designs include asking for a shorter input, returning an error or offering a cloud route. These are hypothetical alternatives. Do not assume that a particular app implements any of them without evidence.
Write down whether the user can see and control that transition. A notice saying that a cloud service will process selected text answers a different question from a general statement that the phone supports local AI. If the behavior is undocumented, the worksheet should preserve that uncertainty.
Read product examples with their date attached
Apple’s September 2025 announcement described Foundation Models access to its on-device model and offline app features on compatible devices with Apple Intelligence enabled. It is a dated example of developers building around local inference, rather than a complete specification of every current Apple AI feature.
The distinction matters because product names persist while models, operating systems and supported routes change. Use a dated announcement to establish what was described at that time. Use current feature documentation to establish the behavior you intend to rely on now.
For the notes worksheet, write the app version and device configuration next to the observation. A feature that works in one configuration is not evidence for every language, device or task. The record can remain small while still being precise.
Follow the output after the model stops
Suppose the fictional action list appears correctly. There are still questions about where it is saved, whether it enters a cloud backup, who can access the note and what happens when a sharing control is used. These questions concern the whole application rather than model inference alone.
Google’s description of AICore says the service isolates requests and does not retain their inputs or outputs after processing. That is a statement about that service. It cannot establish the retention policy of an invented application that receives the output and saves it elsewhere.
Ask for a separate answer about deletion and synchronization. Removing an action list from a screen, removing it from local storage and removing shared copies are distinct outcomes. A clear product explanation should identify which action its control performs.
Keep a compact feature record
Keep one record per feature:
| Field | Question to answer |
|---|---|
| Task | Which exact input becomes which output? |
| Preparation | Which note, model, or resource must already be downloaded? |
| Execution | What documents where this particular request runs? |
| Handoff | When could processing move to another service, and can you control that? |
| Output | Where is the result saved, synchronized, or shared? |
| Open question | Which part remains unverified for this app and version? |
Add actual observations only when you have made them; this article reports no device benchmark or hands-on product test.
For everyday use, that record is more useful than a universal claim that local AI is always private or always faster. It lets you describe the specific benefit you expect and the conditions it depends on. When a feature changes, revisit the relevant row instead of starting again with the marketing label.
