Apple Intelligence is more private than most AI assistants, but not all of it runs on your iPhone. Many requests are handled by Apple’s roughly 3-billion-parameter on-device model. Bigger requests go to Apple’s Private Cloud Compute servers, which Apple built so your data is used only for that request and not stored. And if you choose to send a request to ChatGPT, it goes to OpenAI under OpenAI’s policies, after your phone asks you first.
So “on-device AI” on a modern phone is really three tiers with different privacy. Knowing which tier you’re using tells you where your words go.
The three tiers of phone AI privacy #
| Tier | Where it runs | Who can see it | Examples |
|---|---|---|---|
| On-device | Your phone’s chip | Nobody else | Apple’s on-device model, Gemini Nano, local LLM apps |
| Private cloud | The phone maker’s special servers | Designed so not even the company can | Apple Private Cloud Compute |
| Regular cloud | A company’s data centers | The company, under its policies | ChatGPT, Gemini app, Claude |
How does Apple Intelligence handle your data? #
On the device #
Apple Intelligence runs on the iPhone 15 Pro and later, which have Apple’s A17 Pro chip or newer and 8 GB of RAM, and it needs 7 GB of storage. Its on-device model handles many Writing Tools requests (rewrite, proofread, summarize), notification summaries and other everyday features. Those requests never leave the phone.
Since iOS 26, third-party apps can also use Apple’s on-device model through a Foundation Models framework, which means some app features built on it run locally too.
Private Cloud Compute #
When a request is too big for the phone, Apple can send it to Private Cloud Compute. Apple’s design goals, which it has published in detail, are that your data is used only to answer that request, isn’t stored, isn’t accessible to Apple staff, and that the server software can be inspected by outside researchers. Your iPhone is designed to refuse to connect to servers that aren’t running the published software.
That’s much stronger than a normal cloud service. It still means your request leaves the phone and needs a connection.
iOS 27, due September 14, 2026, leans on Private Cloud Compute more. Apple says its new Siri runs on Private Cloud Compute with a model Apple trained and refined using outputs from Google’s Gemini models, after the two companies announced a partnership in January 2026. Apple says it isn’t running Google’s models directly.
ChatGPT integration #
Apple Intelligence can pass some requests to ChatGPT. Apple says it asks before sending anything, and obscures your IP address. Once a request reaches OpenAI, OpenAI’s own policies apply, and if you sign in with your ChatGPT account, its account settings apply too.
How to see where your requests went #
- Check the prompts. When Siri or Writing Tools wants to use ChatGPT, it asks. You can turn off the ChatGPT extension entirely in the Apple Intelligence settings.
- Look for Apple’s report. Recent iOS versions include an Apple Intelligence Report under Settings > Privacy & Security that can log requests sent to Private Cloud Compute.
- Try airplane mode. Anything that still works with no connection ran on the device.
What about Android and Gemini? #
Android has the same split, with different names.
Gemini Nano is Google’s on-device model. It runs in an Android system service called AICore, and Google says it works “without needing a network connection or sending data to the cloud.” Apps use it for summarizing, proofreading, rewriting, describing images and transcribing speech. It only runs on supported phones, including the Pixel 9, 10 and 11 series, the Galaxy S25 and S26 series and a list of recent flagships from other brands.
The Gemini app is a cloud assistant. Your conversations go to Google, and Google says a subset of them is reviewed by humans. We compare that with other chatbots in do humans read your AI chats.
Samsung’s Galaxy AI mixes on-device and cloud processing. Samsung offers a setting to process data only on the device, which turns off features that need the cloud. Check your phone’s Galaxy AI settings for the current options.
Where fully local AI apps fit #
Built-in phone AI is a set of tools: rewrite this, summarize that, clean up a photo. For an open-ended conversation, you’re usually pointed to a cloud chatbot.
A local AI app fills that gap. It runs an open language model entirely on your phone, with no private cloud and no regular cloud. Personal LLM is one example: it runs models like Qwen 3.5, Gemma 4, Ministral 3 and GLM 4.6V Flash on the phone’s own chip, has no server, and needs no account. The only things that use the network are the one-time model download from Hugging Face and the ads in the free version.
Compared with Apple Intelligence:
| Apple Intelligence | Local LLM app | |
|---|---|---|
| Phones | iPhone 15 Pro and later | iOS 15.1+ or Android 7+ with enough RAM |
| Open-ended chat | Via Siri, with cloud help | Yes, fully on the phone |
| You choose the model | No | Yes |
| Leaves the phone | Sometimes (Private Cloud Compute, ChatGPT) | Never, apart from downloads and ads |
| Works offline | Partly | Yes, after the model download |
| Answer quality | Strong for its tasks, cloud for harder ones | Limited by what fits on the phone |
The honest trade-off is capability. Private Cloud Compute and cloud chatbots can run much bigger models than your phone. A local model is the choice when you’d rather have every word stay on the device, and accept somewhat weaker answers in return.
For more on what each assistant does, see our comparison of Siri, Gemini and local AI assistants, and our guide to built-in AI in iOS and Android.
Frequently asked questions #
Does Apple Intelligence send my data to the cloud? #
Some of it. Many requests run on the iPhone. Larger ones can go to Apple’s Private Cloud Compute, which Apple designed not to store your data. Requests you choose to send to ChatGPT go to OpenAI, and Apple asks before sending them.
Is Private Cloud Compute really private? #
It’s designed to be much more private than normal cloud services. Apple says data is used only for your request, isn’t kept, isn’t accessible to Apple, and that outside researchers can inspect the server software. It still requires sending your request off the phone.
Does Apple Intelligence work offline? #
Parts of it do. Features handled by the on-device model, like many Writing Tools requests, work without a connection. Anything that needs Private Cloud Compute or ChatGPT doesn’t.
Is Gemini Nano private? #
Yes. Gemini Nano runs on the phone, and Google says it works without sending data to the cloud. The Gemini app is different: it’s a cloud assistant, and your chats go to Google.
What’s the most private way to use AI on a phone? #
Run a model locally with an app that has no server and no account, and check it with an airplane mode test. Use cloud services for questions where privacy doesn’t matter.