But what is artificial intelligence that works without internet for?
Part of your phone's AI works inside the phone, part sends your data out. Knowing which does what is the thing that actually matters for your privacy.
- Level
- beginner
- Time
- 10 minutes
- What you need
- Your phone
A new phone arrives with artificial intelligence already inside: it rewrites a message, summarises the notifications you missed, removes a stray passer-by from a photo, transcribes a call. One phrase keeps coming back in the advertising: on-device.
It sounds like a technical detail for insiders. In fact it decides where your sentences, your photos and your voice end up, and it is worth understanding once.
On-device (also "local"): the work happens inside the phone and the data does not leave. In the cloud: the work happens on the company's computers, so what you ask travels over the internet. Model: the program doing the work; small ones fit inside a phone, enormous ones live only in data centres.
Why your phone’s AI is split in two
All else being equal in how it was built, an AI model is more capable the bigger it is, and big models do not fit in a phone: they take far more memory than a phone has, and the work would be more than the battery could carry. Today’s small models are nonetheless much better than the big ones of a few years ago, which is why a phone now handles some of this by itself.
So the companies split the work. Small, frequent jobs go to a small model living in the phone. Jobs that need reasoning, knowledge of the world or long pieces of writing get sent out.
In practice that means when you use “your phone’s AI”, sometimes nothing leaves your device and sometimes everything does. From the outside the two look alike.
What happens inside the phone
What runs locally on recent phones is roughly this: rewriting, correcting or shortening text you have already written; summarising notifications that arrived while you were not looking; searching photos in words and removing an unwanted element from an image; transcribing and summarising a recording or the notes from a call; making small custom images or emoji; and voice recognition for simpler commands.
The sign that a feature works locally is that it keeps working in aeroplane mode, with the connection off. If it still works, the work is happening inside the phone, and that tells you more than anything printed on the box. The reverse does not hold in the same way: a feature can stop working offline simply because it needed to download something, and some decide case by case whether to manage alone or ask outside.
What goes out, and where
Everything else travels over the internet. Actual questions to the assistant, long requests, anything needing current information, and generally anything involving a third-party assistant connected to the phone.
Here there is a distinction worth attention, because it is where companies genuinely differ. Some, when they must send a request out, send it to servers built specifically to keep nothing: the stated promise is that the data is used only to produce the answer and is deleted immediately, never logged, and unreadable by the company’s own staff. In Apple’s case the system is called Private Cloud Compute, and the company makes the software running on those servers available to security researchers, so they can verify that things are as it says. That is a rare level of openness in this industry.
It remains a technical promise, not a law of nature. But it is a checkable promise, and that is very different from silence.
None of these features switches itself on to read your messages behind your back. When an outside assistant is brought into a request, the phone asks you explicitly the first time and you can say no, while carrying on using everything else.
So what is it actually for
Three concrete things, all dull and all useful.
It matters for privacy in the cases that count. Correcting a message where you discuss health, money or another person does not leave the phone. That is a real difference, not marketing.
It matters for speed and battery. Work done locally answers at once and needs no network. Anyone with a shaky connection notices.
It matters because you use it without thinking. Notification summaries and searching photos in words work better precisely because you never decide to “use AI”: they are simply there.
The awkward part: your phone may not have it
These features need memory and a recent processor, and manufacturers cut off models only two or three years old. On Apple phones the line runs between the ordinary iPhone 15 and the 15 Pro, with the non-Pro left out; on the Android side the newest features call for amounts of memory that phones from two years ago often lack.
It is not purely a sales excuse: a model, even a small one, has to sit in memory whole. But the practical consequence is that this is the first time in many years that the gap between a three-year-old phone and a new one shows up in things people use every day, rather than only in the camera.
That is still not a reason to replace a phone now. It is a reason to look at the question when you replace it anyway.
If your phone has these features, use them: they are handy and the most sensitive part stays inside the phone. If it does not, you are not missing much, because everything that genuinely matters can be done with a free app that talks to the internet. What to take away is not a purchase but a question to ask each time: does this thing I am asking stay here, or go out? The aeroplane test answers it in ten seconds.
If you want to control what leaves
In your phone’s settings, under the assistant or artificial intelligence section, there is usually a switch to keep processing on the device and another to turn off connected outside assistants. Switching them off costs you some features, and the choice depends on how much you need them.
If this interests you in the opposite direction, that is, having an AI of your own that talks to nobody, it can be done on a computer too: I cover it in what an AI running on your own computer is for. And if you want to know why these things sometimes get it wrong so confidently, the explanation is in what a language model is.