Apple Intelligence privacy starts with where the request is processed

The phrase on-device AI is useful only when it describes the particular task being performed. Apple Intelligence includes processing on the device and, for some requests, a server-based path called Private Cloud Compute. A privacy discussion should distinguish those paths.

What Apple describes

Apple’s guide says many Apple Intelligence tasks run on the device. More demanding requests can use Private Cloud Compute, with relevant information processed on Apple silicon servers. Apple states that this data is not stored or made accessible to the company and describes independent inspection of the system.

These are Apple’s documented design and privacy claims. This article is an explanation of that documentation, not an independent security audit or a guarantee about every app that uses AI.

Ask a more precise question

Instead of asking whether a product is private in the abstract, ask what information a particular feature uses, where processing occurs, and which organization’s terms apply. A feature connected to an external service may have a different boundary from an Apple-operated feature.

Our suggested review is task by task. Consider a writing request, an image task, and an action involving personal context separately. The fact that one operation can run locally does not establish how every other operation works.

Keep sensitive decisions deliberate

For work information, follow the organization’s rules about approved services and data handling. For personal information, share only what is needed for the task and review the available controls. Avoid pasting a complete document when a small, non-sensitive excerpt would answer the question.

Privacy and accuracy are also separate questions. A well-protected processing path does not prove that the generated result is correct. Review important output before acting on it, especially when it affects another person or changes information in an app.

The useful takeaway is a clearer map of the system: identify the feature, the data it needs, the processing path, and the terms that govern it. That map is more informative than assuming every AI request stays on the phone or that every cloud request is handled identically.