On-device AI is AI that does its computing inside your own phone or laptop, not on a server. It uses small models that fit in the device's memory, so it can work offline. How fast and capable it is depends on the model and the device.
A question from the app. Pick an answer.
What do you need to use it offline?
A cloud account does not replace the model and software that run on your device.
You need compatible hardware and software that runs the model without remote services.
Generating locally without remote services does not require a daily connection.
Many short tasks already run on the phone: dictating a message, suggesting the next word on the keyboard, recognizing faces in your photos or summarizing a notification. They are small, frequent jobs, exactly what a small model does well.
Heavy work usually goes to the cloud: a long conversation with an AI chat, generating an image or analyzing a huge document. Many apps mix both: they try to solve it on the phone and, if that is not enough, send the request to a server.
Upsides: your data can stay on the phone, it can work on a plane or with no signal, it answers without waiting for the network and there is no per request cost for anyone to pay. That is why, on many phones, dictation keeps working when you run out of data.
Limits: the model has to fit in the phone's memory alongside your other apps, so the device's memory and power limit it. Computing drains the battery and warms up the device, and older phones sometimes lack these features.
A clue: turn off Wi-Fi and mobile data and try it with something new. If it keeps working, the computing is probably local. To confirm, read the app's privacy settings or documentation: many spell out what is processed on the device and what is not.
Both run on your own device, but you use them differently. A local model is one you pick, download and install on your computer, and you can swap it for another. On-device AI usually comes built into the system or an app: you do not see it or choose it, it just works.