Almost every new Android phone now ships with an AI badge somewhere on the box. The phrase hides an important split: some of those features run entirely on the handset, and some quietly send your request to a data centre. Knowing which is which changes what you should trust them with.
What actually runs on the phone
On-device AI is handled by the phone’s NPU, a small, power-efficient chip built for matrix maths. It is good at narrow, repeated jobs where the model is small enough to stay in memory:
- Live transcription, captioning and dictation
- Photo cleanup: object removal, upscaling, noise reduction
- Call screening and noise suppression
- Short summaries of messages, notes and recordings in some apps
Because the model never leaves the device, these features keep working in airplane mode and answer in milliseconds.
What still goes to a server
Long-context reasoning, image generation, and anything that has to reason about the wider web are usually too large to fit a phone’s memory budget. Those requests are routed to cloud models, where the prompt, and often the attached image or document, is processed remotely.
Why the distinction is worth caring about
- Privacy: on-device work stays on the device; cloud work is governed by the provider’s retention policy.
- Latency: local inference never waits on a network round-trip.
- Cost and battery: NPU work is cheap per query; long cloud calls drain battery through the radio.
- Reliability: only the local half survives a weak signal.
How to tell which one you are using
There is no universal settings switch, but three habits expose the answer quickly:
- Turn on airplane mode and repeat the action. If it still works, it ran locally.
- Watch the data counter before and after the feature runs. A jump of megabytes means a server was involved.
- Read the feature’s own settings page. Vendors increasingly label options as on-device or in the cloud, sometimes with a separate toggle.
The practical takeaway
Treat on-device AI as a fast, private assistant for small, well-defined jobs, and treat cloud AI as a more capable service that needs a moment, and a decision, before you hand it something sensitive. Neither is better in the abstract. The mistake is assuming every AI feature on your phone behaves the same way.
