AI on your phone: what runs on the device and what gets sent out
«Processed on your device» is the phrase of the moment in phone adverts, and it is not a lie. It is not the whole truth either: in practice there is a split, the most interesting part usually needs the cloud, and the user rarely sees where the border sits.
Manufacturers have spent several years repeating the same promise: the artificial intelligence in your phone runs inside your phone, so your data never leaves it. That is true of a substantial part of it, and it is why you can dictate a message with no signal, or search your gallery for “dog” and get the dog.
But there is a split, and it is worth understanding where the line falls.
What genuinely stays on the device
Modern phones carry a processor dedicated to this kind of maths. It is small and efficient, and it is enough for lightweight models handling specific jobs:
- Dictation and transcription. Turning speech into text.
- Autocorrect and keyboard suggestions.
- Face and object recognition in the gallery, so you can search for “beach” or group by person.
- Text inside images, so you can copy or translate it.
- Suggested quick replies in messages.
- Call screening and spam filtering.
They share two traits: they are frequent and they are small. Shipping them to a server would be slow, expensive and a constant drain on your data allowance. Running them locally makes technical sense before it makes ethical sense — privacy, here, is a pleasant consequence of an engineering decision.
What almost always leaves
The moment a task needs a large model, the phone cannot cope. A model capable of drafting a whole email takes far more memory than a handset has and would flatten the battery in minutes. So the request goes to a server:
- Writing, rewriting or expanding text.
- Summarising a document or a web page.
- Answering general assistant questions.
- Generating or creatively editing images.
- Long, high-quality translation.
Here it is the content that travels: the text of the email, the document, the photo. And once it is out, everything in the piece on what free AI services do with your data applies.
The grey zone: automatic routing
The awkward part is that many systems decide on their own where to run each request. If it is simple it stays in; if not, it goes out. The user phrases the question the same way either way and does not always see the difference.
Some platforms have added intermediate schemes where the request is handled on their own servers under reinforced guarantees and without being tied to your identity. That is a genuine improvement over handing it to any third party, but it should not be confused with local processing: the data still leaves the device.
Practical trick: put the phone in aeroplane mode and try the feature. If it still works, it runs locally. If it asks for a connection, it goes out.
It is not a perfect test — some features simply refuse to work offline for other reasons — but as a kitchen-table check it is surprisingly reliable.
The keyboard, which sees the most
It deserves its own section because it is the app that the most sensitive information passes through: passwords typed by hand, private messages, searches.
System keyboards handle prediction locally. Third-party keyboards installed separately are another matter: to work they need “full access” permission, and from there what they do depends entirely on their maker. If you are going to install one, make it a company you can name.
It is the same test you apply to browser extensions and, at bottom, to any middleman: whoever sits in the middle sees everything.
What to check in settings
On Android, inside your Google account settings, you will find the activity controls for the assistant and the AI apps: you can review, delete and schedule automatic deletion.
On iPhone, the Siri and AI feature settings let you decide what is shared and clear the request history.
In both cases, two questions are enough:
- Is it keeping a history of my requests?
- Can I delete it and have it delete itself from now on?
If the answer to the second is yes, set it to the shortest window you can live with. It is the same auto-deletion reasoning covered in how to delete your ChatGPT history.
What you can do today, in five minutes
- Test your most-used AI features in aeroplane mode, so you know which ones leave.
- Review the assistant’s activity controls and set automatic deletion.
- Check which keyboard you have installed and who makes it.
- For anything genuinely sensitive, use the local app rather than the cloud assistant.
The short version
- Small and frequent — dictation, keyboard, gallery — runs on the phone.
- Anything needing a large model — writing, summarising, generating — goes to the cloud.
- Many systems choose for themselves, and the user cannot see the border.
- Aeroplane mode is the kitchen-table test for where something runs.
- A third-party keyboard with full access sees everything you type.
Frequently asked
The questions that keep coming up
What does on-device AI actually mean?
That the model runs on the phone's own processor and the content never leaves for the internet. It is used for small, frequent tasks: dictation, autocorrect, suggested replies, finding objects in your photos.
Which AI tasks do go to the cloud?
The ones that need a large model: writing long text, summarising documents, answering complex questions or generating images. A phone has neither the memory nor the power budget to run those.
Can I tell when something is sent?
Not always precisely. Some systems warn you when a request needs a connection, and others say so in their documentation. The most reliable indirect signal is whether the feature stops working in aeroplane mode.
Are photos analysed on the phone or on a server?
Face and object recognition for organising your gallery is normally done on the phone itself. Advanced generative editing usually needs servers.
