What you need to know
On-device AI performs model inference on the phone, tablet, or computer in front of you. A prompt can be processed without being sent to a hosted model, but downloads, updates, web search, syncing, and reports may still need a network. Judge each feature by its actual data flow, not by a product label.
Start with the location of inference
The useful dividing line is where the model evaluates your prompt. In a local system, the model weights and runtime are stored on your hardware and the computation happens there. In a cloud system, an app sends a request to remote infrastructure and receives the generated result. Some products combine both approaches, so the answer can change from one feature to another.
Local execution can reduce routine disclosure to an inference provider and keep core tasks available without connectivity. It also moves practical constraints onto your device. Memory, storage, battery, heat, runtime compatibility, and model size all affect what is usable. Larger cloud systems may handle difficult work better, while a compact local model can be the more predictable choice for private or disconnected tasks.
Verify the boundary instead of trusting a slogan
Ask whether the exact feature you plan to use performs inference locally. Then separate normal chat from model downloads, online search, telemetry, crash reports, backups, and synchronisation. An app can offer genuinely local chat while using the internet for clearly identified optional actions.
A repeatable test is more valuable than a broad claim. Download the required model, begin a new conversation with Wi-Fi and mobile data disabled, and check that fresh output is generated. Pair that test with the product documentation and privacy notice because airplane mode cannot reveal earlier uploads, remote retention, or what happens after reconnection.