Mostly, yes. When an AI model runs on your phone, your prompts, photos and answers are computed and stored on the device, so there’s no server log, no retention policy and nobody on the other end reviewing your chats. Deleting a conversation actually deletes it, because there was no second copy.
What “on-device” doesn’t cover is everything around the model: apps that fall back to the cloud on hard questions, ad and analytics code inside the app, your keyboard, device backups, and anyone who can unlock your phone. This is what each of those exposes, and what a clean on-device design looks like.
What does on-device AI actually protect? #
With a cloud chatbot, every message is sent to the provider, processed on its servers and stored under its policy. That creates exposure you can’t control: breaches, legal requests, staff or contractor review, use for training, and retention after you hit delete.
When the model runs locally, those risks go away because the data never leaves. In practice:
- Your prompts and the model’s answers are generated by your phone’s own chip.
- Photos and documents you ask about are read on the device.
- Chat history sits in the app’s storage on the phone.
- Deleting a chat removes the only copy.
What can still leave your phone? #
“On-device” describes where the model runs, not everything the app does. These are the leak paths worth knowing about.
| Leak path | What it can expose | How to check |
|---|---|---|
| Cloud fallback in “hybrid” apps | The content of some requests | Read the app’s docs; test in airplane mode |
| Model downloads | That you downloaded a model, plus your IP address | One-time; unavoidable unless you sideload the file |
| Ad and analytics SDKs | Device information and app usage, not your chats | The store’s privacy label |
| Cloud backup or sync | Chat history, if the app syncs or your backup includes it | The app’s settings and your backup settings |
| Your keyboard | Everything you type, if a third-party keyboard has network access | Keyboard settings |
| Sharing and screenshots | Whatever you send on | Your own habits |
| Physical access | Everything on the phone | Passcode, auto-lock, biometrics |
The keyboard one surprises people. A model can run entirely on the device while a cloud-connected third-party keyboard sees every word you type into it. On iPhone that’s the “Full Access” permission a keyboard asks for; on Android it’s whatever network access the keyboard app has.
What does a local AI app’s network traffic look like? #
As a worked example, here’s the full network ledger for Personal LLM, taken from its published documentation.
| What | Goes over the network? | Details |
|---|---|---|
| Model files | Yes, once | Downloaded straight from Hugging Face; pausable and resumable |
| Ads in the free version | Yes, only while online | Google AdMob; on iOS the app asks for tracking permission, and declining means non-personalized ads |
| Your chats, prompts and photos | Never | Processed on the phone’s chip; there is no Personal LLM server |
| Settings and system prompts | Never | Stored on the phone |
| Usage stats | Never | Kept on the phone |
| Backups | Only if you move the file | Exported as a single plain JSON file you control, not an encrypted archive |
There’s no account, no cloud sync and no analytics on conversations. Uninstall it and everything is gone, because nothing was ever stored anywhere else. The ads are the honest asterisk: while you’re online, a free ad-supported app is making network requests, even though the ad network never sees a single word of your chats. A one-time purchase removes them.
That’s the shape of a clean on-device design. The model comes in once; your data never goes out.
How is this different from “private cloud” AI? #
Three tiers get described with similar language, and they’re not the same thing.
- Fully local. The model file is on your phone and the answer is computed there. Works in airplane mode.
- Private cloud. Your request goes to a company’s servers with extra protections around it, such as Apple’s Private Cloud Compute. Better than ordinary cloud processing, but your text still leaves the phone.
- Ordinary cloud. Your request goes to the provider, is stored under its retention policy, and may be reviewed or used for training.
An app can mix them. Apple Intelligence handles many requests on the phone, sends heavier ones to Private Cloud Compute, and passes others to ChatGPT if you enable that extension. Assuming one tier when you’re getting another is the mistake worth avoiding.
Does on-device AI help with privacy laws? #
Privacy laws such as the GDPR and California’s CCPA mostly regulate what organizations do with personal data they collect. If a company never receives your prompts, there’s much less for it to collect, store, secure or disclose, which is why on-device processing gets described as privacy by design.
It isn’t an exemption. An app can run its model locally and still collect personal data through analytics, ads or accounts, and those parts are regulated as usual. This is general background rather than legal advice, and nothing here says any particular app meets any particular regulation. If you’re handling other people’s information at work, using AI with confidential client data covers the professional rules to check.
How do you confirm an app is what it says? #
The short version is the airplane-mode test: download whatever the app needs, switch on airplane mode with Wi-Fi off, reopen the app and ask it something. If it answers, the model is local. Then read the store privacy label for what it collects when you’re online. How to tell if an AI app is really private walks through all ten checks, and AI app privacy labels explained covers reading the label itself.
Once you’ve picked an app, the risks that remain are device-side ones: your lock screen, your backups, and where your exported files end up. Are offline AI apps safe? goes through those.
Frequently asked questions #
Is on-device AI completely private? #
The model’s processing is, since your prompts and photos never leave the phone. The app around it may still use the network for model downloads, ads or analytics, and your phone’s backups and lock screen still matter. Check the privacy label and try the app in airplane mode.
Can an offline AI app still track me? #
It can if it includes advertising or analytics code, which collects device and usage data while you’re online. That’s separate from your conversations. On iPhone you can decline tracking when asked, and ads then stay non-personalized.
Does deleting a chat really delete it? #
With an on-device app, deleting a chat removes the only copy, apart from any device backup that already captured it. With a cloud chatbot, deletion follows the provider’s policy, which can include a retention window and exceptions for reviewed or legally held data.
Is “on-device AI” the same as a private cloud? #
No. On-device means the model runs on your phone and nothing is sent. A private cloud means your request goes to a company’s servers with extra protections. Both are better than ordinary cloud processing, but only one keeps your text on the phone.