Ask five Android owners what AI their phone has and you will get five answers, because Android no longer has a single AI layer. There are at least four, and they are controlled by different companies.
Layer one: the operating system
Android itself provides the plumbing: on-device inference APIs, transcription hooks, and the permissions model that decides which app may use them. This layer is mostly invisible. It matters because it determines whether a feature can run offline at all.
Layer two: the assistant
The voice assistant is the layer users actually notice. Depending on the device, it may be the manufacturer’s assistant, a third-party model, or a hybrid that routes simple commands locally and harder questions to a server. The routing is rarely documented, which is why the airplane-mode test remains the fastest way to find out.
Layer three: the manufacturer’s additions
Summarising calls, cleaning up photos, translating subtitles, screening unknown numbers. These are the features printed on the box, and they are usually the first to disappear when a device stops receiving updates. They are also the hardest to audit, because they live in system apps rather than in settings you can inspect.
Layer four: your apps
Everything else is an app making its own choice. Two apps using the same model may behave differently because one sends your data to a server and the other does not. The app’s own privacy page is the only reliable source.
What you can control
- Turn off cloud processing per feature where the toggle exists.
- Revoke microphone and notification access from assistants you do not use.
- Check whether a feature survives airplane mode before trusting it with sensitive content.
- Prefer apps that state whether inference is local or remote.
Bottom line: treat Android AI as four separate products with four separate privacy stories. The label on the box describes the second one only.
