Guide

What 'Data Not Collected' Actually Means — and How to Verify Any AI App's Privacy

If you're looking for an AI app that doesn't collect your data, you have a problem: every app says it respects your privacy. Marketing copy is free. What you need is a claim you can check yourself, in under a minute, without trusting anyone's word — including this page's.

The App Store gives you exactly that. Every app's product page carries a privacy label the developer is required to declare, and one label — "Data Not Collected" — is the strictest claim an app can make. This guide explains what that label means, how to read it on any app in about 30 seconds, and which additional checks catch what a label can't. The method works on any app, including competitors of ours. That's the point.

The App Store privacy label, explained

Since late 2020, Apple has required every app submitted to the App Store to declare its data practices. The result is the App Privacy section on each app's product page — often called a privacy "nutrition label." Developers must state, category by category, what data the app and its third-party partners collect: contact information, location, identifiers, purchase history, browsing history, usage data, diagnostics, user content, and more.

Each declared category lands under one of three headings, from worst to best for you:

• Data Used to Track You — data shared with data brokers or used to follow you across other companies' apps and websites, typically for advertising.
• Data Linked to You — data collected and tied to your identity (through an account, a device identifier, or similar).
• Data Not Linked to You — data collected, but which the developer says is not tied to your identity, such as anonymized analytics.

One detail matters enormously for AI apps: in Apple's framework, data counts as "collected" when it is transmitted off your device in a way the developer or their partners can access. Data that is processed entirely on your device and never transmitted is not collected at all. That distinction is the entire architecture argument this guide builds toward — our guide to on-device AI on iPhone, iPad, and Mac covers what local processing can and can't do.

'Data Not Collected': the strictest label an app can carry

When an app's App Privacy section shows a single entry — "Data Not Collected" — the developer is declaring that neither they nor any third party integrated into the app collects any data from it. Not anonymized analytics. Not crash diagnostics tied to your device. Not usage statistics. Nothing in any category.

This label is rare among AI apps, for a structural reason: most AI apps are thin clients in front of a cloud model. Your question — and often your documents — must be transmitted to a server to be processed at all. That's collection by definition, before any analytics SDK even enters the picture. An AI app can only truthfully carry "Data Not Collected" if the model inference itself happens on your device.

"Data Not Collected" is not a promise to handle your data carefully. It is a declaration that there is no data leaving the app to handle.

How to check any app's privacy label in 30 seconds

1. Open the app's page in the App Store (or its apps.apple.com listing in any browser — no iPhone required).
2. Scroll down to the App Privacy section, below the ratings and description.
3. Read the headings. "Data Not Collected" appears alone if it applies. Otherwise, tap See Details to see exactly which data types sit under "Data Used to Track You," "Data Linked to You," and "Data Not Linked to You."

What to look for, in order of importance: anything under Data Used to Track You is the sharpest red flag for a privacy-sensitive use case. Data Linked to You containing "User Content" means your actual documents or conversations are collected and tied to your identity. A long list under any heading in an app that claims to be "private" tells you the marketing and the declaration disagree — believe the declaration, since misdeclaring it violates App Store rules while marketing copy carries no such obligation.

Policy vs. architecture: a promise not to misuse data vs. no data to misuse

A privacy policy is a promise about behavior: "we collect your data, but we won't sell it," or "we delete it after 30 days." Promises have failure modes. Policies get rewritten after acquisitions. Terms change with a notification email nobody reads. Employees make mistakes. Servers get breached, and a breach exposes whatever was collected — regardless of what the policy promised.

Architecture is a different kind of claim. If an app never transmits your data anywhere, there is nothing to sell, nothing to leak, nothing to subpoena, and nothing for a breach to expose. The question shifts from "will they misuse my data?" to "do they ever have my data?" — and only the second question has a verifiable answer. Can't is stronger than won't.

This distinction is why "we don't train on your data" — the standard reassurance from cloud AI services — answers the wrong question for anyone working with sensitive material. Your documents still crossed the network and sat on someone else's infrastructure. If that's the scenario you're worried about, we've written a full guide to whether it's safe to upload confidential documents to cloud AI.

Three more checks anyone can run

The privacy label is a declaration. These three checks are observations — you're testing the app's actual behavior, not reading its paperwork.

1. The airplane-mode test

Turn on airplane mode and use the app's core feature. If it keeps working with every radio off, the processing demonstrably happens on your device — there is no server available to send anything to. If it throws a connection error, it's a cloud app, whatever the marketing says. One honest caveat: this proves the offline session stayed local, not that the app never syncs anything once you reconnect. It's a strong signal, not a formal proof — which is why you stack it with the label and the other checks rather than relying on it alone. We use this same test in Does Apple Intelligence work offline?

2. The account test

Does the app work without creating an account? An account means a server-side record with, at minimum, an identifier tied to you — and usually usage history attached to it. An app with no sign-up, no sign-in, and no email field has no user database to breach and no profile to build.

3. The API-key test

Some "private" AI apps ask you to paste in an API key from a cloud AI provider. That's an admission of architecture: the app is a front end, and your content is transmitted to whichever service issued the key, under that service's terms — not the app's. An app that never asks for an API key and never offers a field for one isn't routing your content to a third-party model.

A worked example you can verify right now

Full disclosure: this site belongs to OpenIntelligence, an AI document app. So don't take our word for anything — run the checks from this guide against us, right now:

1. The label. Open OpenIntelligence on the App Store and scroll to App Privacy. The label reads "Data Not Collected." You're reading Apple's page, not ours.
2. The account test. Download it free and open it. There is no account to create, no sign-in screen, and no email field — nothing to attach a server-side profile to.
3. The API-key test. There is no API-key field anywhere in the app. It runs on Apple's on-device foundation models, not on a cloud service you'd need credentials for.
4. The airplane-mode test. Import a PDF, turn on airplane mode, and ask a question about it. It keeps answering — with tappable citations into the source — because retrieval, verification, and answering all run on the device.

What the app actually does: it ingests PDFs, Office and iWork documents, spreadsheets, presentations, notes, scans, code, and audio or video recordings; indexes them locally; and answers questions with citations you can tap to see the exact source passage. When your documents don't support an answer, it says so and abstains instead of guessing. The retrieval engine itself is open source, so the architecture can be inspected, not just tested from the outside.

One requirement stated plainly, because the on-device approach depends on the hardware: OpenIntelligence needs iOS, iPadOS, or macOS 26 or later and Apple Intelligence-capable hardware — an iPhone 15 Pro or later, or an iPad or Mac with an M1 chip or later. On older devices it won't run, and no cloud fallback exists to quietly change the privacy story. The app is free to download; a one-time Lifetime unlock and Pro subscriptions remove the limits.

What the label doesn't tell you

Honesty requires the last section to cut against the previous ones. The privacy label has real limits:

• It's self-reported. Developers fill in the declaration themselves. Apple requires it to be accurate and can act against apps that misdeclare, but it does not independently verify every label before publishing it. Researchers and journalists have repeatedly found apps whose observed network behavior didn't match their labels.
• It has scope exclusions. Apple's rules allow certain optional, limited disclosures to be omitted in narrow cases, and the label describes collection, not security quality — a "Data Linked to You" app might handle your data impeccably, and a mislabeled app might not.
• It's a snapshot. Labels can change with updates. The label you verified last year is not evidence about the version you're running today.

This is precisely why the architectural checks matter: they don't ask you to trust a declaration. Airplane mode either works or it doesn't. An account field either exists or it doesn't. Each check is weak alone and strong stacked — a truthful "Data Not Collected" label, a core feature that works offline, no account, and no API key describe an app that has no channel for your data to leave through.

We built OpenIntelligence to answer only what its sources support and to abstain otherwise. This page is held to the same standard: every claim in it about our app is one you can check yourself in under a minute, and the ones you can't check, we've told you so.