Privacy and AI

On the device or in the cloud: where a to-do app’s AI runs

A to-do app’s AI runs in one of three places. On your iPhone, using Apple’s on-device model, nothing leaves the phone and it works offline, but the model is small. In Apple’s Private Cloud Compute, which Apple says neither stores your request nor makes it accessible to Apple. Or with a provider such as OpenAI, often through the app’s own server, where both their policies decide what is kept. Short, well-defined jobs suit the device; long conversations usually need the cloud.

Coming soon to iPhone and iPad

Last checked

On this page

Three places an AI request can run

Where it runs What leaves your iPhone Works offline Suits
On the device, with Apple’s model Nothing Yes Short, well-defined jobs: sorting, short suggestions, tidy wording
Apple’s Private Cloud Compute The request, encrypted to Apple’s servers No Larger Apple Intelligence requests
A provider such as OpenAI, often through the app’s server The request and whatever context the app adds No Long conversations, broad knowledge, many steps

The rest of this page goes through each, and what you can check.

On your iPhone

Since iOS 26, apps can use Apple’s Foundation Models framework to send requests to the on-device language model at the core of Apple Intelligence. Apple describes it as a 3 billion parameter model, and says features built on it protect privacy, work offline and cost nothing per request.

That is the strongest privacy position an AI feature can have: the text never leaves the phone, so there is no server policy to read. It has two limits:

  1. Size. A model that fits on a phone is good at narrow tasks with a clear shape, such as turning a sentence into a few steps or sorting items into groups. It is weaker at long, open conversations and at questions that need wide world knowledge.
  2. Hardware. Because the framework uses the model at the core of Apple Intelligence, it depends on a device that supports Apple Intelligence. Apple’s list for iOS 27 includes iPhone 15 Pro and 15 Pro Max, iPhone 16 models and later, and iPhone Air, with the device and Siri language set to the same supported language, and several gigabytes free for the models (Apple says up to 8 GB on many models, up to 14 GB on the newest).

So an app that uses the on-device model has to decide what happens on an older phone: the feature is missing, or it falls back to the cloud. That is worth asking about.

Apple’s Private Cloud Compute

When an Apple Intelligence task needs more than the phone can do, Apple says the request may go to Private Cloud Compute (PCC). Apple’s privacy page says the data “is not stored or made accessible to Apple”, is processed only to fulfil your request, and is not retained once the result comes back.

Apple’s security team published the design in June 2024 as five requirements:

  • Stateless computation. Personal data is used only for the request and not kept, including in logs.
  • Enforceable guarantees. The promises have to be enforced by the system’s design, not by policy alone.
  • No privileged runtime access. Not even Apple’s site reliability staff get a back door during an outage.
  • Non-targetability. An attacker should not be able to aim at one person’s data without attacking the whole system.
  • Verifiable transparency. Apple publishes the software images so security researchers can check them, and devices only send requests to servers that prove they run publicly listed software.

There are honest limits. The EFF points out that nothing on screen tells you whether a request ran on the device or in PCC, and that “private” here means engineered so Apple should not see or keep the data, not that the data stays on your phone. Apple also says some features that rely on server-side models have daily usage limits. If you want to see what went out, Apple describes an Apple Intelligence & PCC Report under Settings, Privacy & Security.

PCC is for Apple’s own features. A third-party app that uses its own provider is not covered by these promises.

A third-party provider

Many to-do and notes apps send assistant requests to a company such as OpenAI, either directly or through a server the developer runs. Here two sets of rules apply: the developer’s, for its server, and the provider’s. OpenAI says data sent to its API has not been used to train its models since 1 March 2023 unless the customer opts in, and that abuse monitoring logs, which can include prompts and responses, are kept for up to 30 days by default, with a Zero Data Retention option for approved customers.

This is where you need the app to be specific. Ask:

  1. Which provider, by name?
  2. Does the request pass through the developer’s server, and does that server keep anything?
  3. What does the request contain: your message only, or items from your list too? Our guide to what an AI assistant sees in your to-do list explains why this varies.
  4. Does it ask before the first request, and can you keep some items out? See how to keep notes and to-dos away from AI.

Which to choose

For most people the sensible rule is: let small, frequent jobs run on the device, and use the cloud on purpose, for things that need it. A feature that sorts your list or suggests a few steps can usually stay on the phone. A conversation about your whole week usually cannot, and that is fine if you know where it goes and you have kept out what should stay private.

Where Dozen helps, and where it doesn’t

Dozen splits its AI the way this page describes. Suggest steps uses your iPhone’s own model, and asks OpenAI only with Dozen Pro, after your permission, when the device cannot make the steps itself. Ask, in Quick Find, lists the notes and to-dos that match a question and, when your iPhone can, writes a short answer above them. The Assistant is optional, and it asks first: nothing is sent to an AI provider until you allow it. Cloud requests go through Loopware’s own server to OpenAI, which passes each one on without logging or storing what you wrote. Anything you tag No AI is kept out of the Assistant’s requests. There is more on Search and Ask and the Assistant.

Where it doesn’t help: Everything in Dozen except the Assistant’s cloud answers works offline. So the Assistant’s cloud answers need a connection, and they do not run on the device. The Assistant is part of Dozen Pro. Features that use the iPhone’s own model depend on a phone that supports Apple Intelligence. And Dozen is not on the App Store yet.

Questions and answers

Is Apple Intelligence processed on my iPhone?

Often, but not always. Apple says many Apple Intelligence models run entirely on the device, and that requests needing a larger model may go to Private Cloud Compute. The EFF notes there is no on-screen sign of which one handled a given request; Apple’s Apple Intelligence & PCC Report setting lists the ones sent to Private Cloud Compute.

Which iPhones can run Apple’s on-device model?

Apple lists iPhone 15 Pro and iPhone 15 Pro Max, iPhone 16 models and later, and iPhone Air, among others, for Apple Intelligence in iOS 27. Device and Siri language must match a supported language, and the models need several gigabytes of storage.

Is cloud AI always less private than on-device AI?

Not always, but on-device is the simpler promise: nothing leaves the phone. A cloud request can be well protected, as Apple describes for Private Cloud Compute, or kept briefly by a provider, as OpenAI describes for its API. In the cloud you are relying on policies and engineering you cannot check yourself.

Sources

What this page says about Dozen was checked against the app on .

Last checked:

Get Dozen

Dozen is free to download. Dozen Pro, a monthly or yearly subscription, adds the Assistant.

Coming soon to iPhone and iPad

Dozen: Tasks & Notes

Dozen is free to download. Dozen Pro, a monthly or yearly subscription, adds the Assistant.

Coming soon to iPhone and iPad