What on-device processing means for your privacy
Phones and laptops now run some of the work that used to happen on a company's servers. Here is what changes, and what does not.
When an app answers a question, corrects a photo or transcribes a voice note, that work runs in one of two places: on the device in your hand, or on a computer somewhere else. Which one it is decides who can see the input.
Where the work happens
Sending work to a server has been the default for years because servers are bigger. The trade is that whatever you send has to leave the device to get there. Running the same work on the device removes that trip, which is why companies talk about it: it is a genuine privacy difference, not only a speed one.
The catch is that a phone has a fraction of a server's capacity. On-device work therefore tends to cover the narrow, frequent jobs, and the heavier ones still go out.
What stays on the device
Typically: keyboard prediction, face and object grouping in your photo library, wake-word detection, live captions, and short transcription. These are small models, tuned for one job, and they run without a connection. If a feature keeps working in airplane mode, that is a reliable sign it is local.
What still leaves
Long answers, image generation, search, translation of anything substantial, and most assistant conversations still go to a server. Some products split the job: the device decides whether it can answer, and passes the rest along. That split is usually invisible, which is the part worth paying attention to.
Local processing also does not mean nothing is recorded. An app can run a model on your device and still send the result, or a log of it, somewhere else. On-device is about where the computation happens, not about what the app does afterwards.
How to check
- Turn the connection off and see which features survive.
- Read the feature's own settings page, not the marketing page. The distinction is usually stated there.
- Look for a setting that controls whether your inputs are used to improve the product. That is a separate question from where the model runs.
On-device processing is a real improvement. It is not a guarantee, and no single label tells you everything. Checking takes a minute.