Key takeaways
- “On-device” should name a specific processing step, not imply that the entire app is permanently offline.
- Local inference, local storage, optional iCloud, App Store verification, and user-initiated sharing are separate data paths.
- The most credible privacy claim is one you can test through permissions, offline behavior, exports, and product-specific disclosures.
The short answer
An iPhone AI app is meaningfully on-device when it identifies the exact task that runs locally, such as transcription, summarization, OCR, search, or classification, and that task still works without sending the input to the developer’s server. That does not prove every feature is offline. You still need separate answers about storage, model downloads, purchases, iCloud, third-party services, exports, diagnostics, and support.
“Private AI” and “on-device AI” are useful terms only when they describe an inspectable data path. A logo, lock icon, or broad promise cannot tell you whether a recording is uploaded for speech recognition, whether a journal entry is sent elsewhere for a summary, or whether an optional sync service creates another copy. The right comparison begins with what the app does to one real piece of data from input to result.
What on-device AI actually means
Apple’s Foundation Models framework gives developers access to an on-device language model for tasks such as summarization, entity extraction, text understanding, refinement, and structured output. Apple also provides local frameworks for speech, Vision OCR, image analysis, Natural Language processing, and other focused work. When an app uses those capabilities directly on supported hardware, the private input can remain inside the device boundary for that named task.
Local inference and local storage are different questions
An app can process a prompt locally and still store the result in a synced database. It can also store everything locally while sending one optional request to a remote service. Ask where the original input lives, where derived data lives, whether a device backup contains it, whether iCloud is optional, and whether deleting the app removes every copy. “The AI runs locally” answers only the inference question.
| Data path | What it can mean | What to verify |
|---|---|---|
| Local inference | The model processes the input on the iPhone, iPad, or Mac. | Which exact features use it, supported hardware, fallback behavior, and whether the input is ever uploaded. |
| Local storage | Records remain in the app sandbox or an App Group container. | Encryption, device backups, deletion, export, and whether extensions can read the same store. |
| Optional iCloud | Selected records may sync through the user’s Apple account. | Whether it is opt-in, which data types sync, whether media files are included, and what happens after disabling it. |
| Apple services | WeatherKit, HealthKit, StoreKit, or another system service may be used. | Permission scope, data sent to Apple, local caching, and whether the feature works when permission is denied. |
| Third-party service | A connected provider may supply bank data, collaboration, or another optional capability. | Provider name, authentication flow, fields shared, retention, disconnect behavior, and a non-connected alternative. |
| User-initiated export | You create a portable copy and choose a destination. | Format, included fields, excluded private history, and the security of the destination after export. |
Scroll horizontally to read the complete comparison on smaller screens.
Seven checks you can perform before trusting the claim
- READ THE PRODUCT-SPECIFIC PRIVACY PAGE: Look for concrete nouns such as audio, transcript, entry, photo, location, and transaction, not only adjectives such as secure or private.
- CHECK THE APP STORE PRIVACY LABEL: Apple requires developers to disclose data collected by them and their partners, while noting that data processed only on the device is not considered collected.
- REVIEW PERMISSIONS IN CONTEXT: A microphone, photo, contacts, location, or Health permission should appear when its feature is used and explain why it is needed.
- TEST AIRPLANE MODE AFTER SETUP: Try the core workflow from input through result and export. Record which steps work and which clearly request a connection.
- LOOK FOR A FALLBACK: A responsible local-AI product explains what happens on unsupported hardware, when the model is unavailable, or when a permission is declined.
- EXPORT AND DELETE A TEST RECORD: Confirm that a useful format is available and that deletion behavior is understandable without contacting support.
- IDENTIFY EVERY OPTIONAL CONNECTION: Model downloads, StoreKit, WeatherKit, iCloud, Plaid, web links, and support can be legitimate while still needing disclosure.
Why Obsidian Ridge Labs focuses on Apple platforms
A narrow platform strategy is an engineering constraint, not a claim that Apple makes every app private. Supporting one hardware and software family gives a small studio more time to test permission timing, local model availability, offline failure states, accessibility, exports, and optional connections. It also lets each product publish specific operating-system and hardware requirements rather than hiding behind a vague compatibility promise.
The platform still does not choose the app’s business model, third-party SDKs, storage design, or support workflow. Those remain developer decisions. Our standard is therefore product-specific: name what stays local, name what connects, label unreleased behavior as provisional, and correct a claim when the implementation does not support it.
Questions people ask about private iPhone AI apps
Questions, answered plainly
Does on-device AI mean an app never uses the internet?
No. It means the named AI processing task runs locally. Downloads, purchase verification, optional iCloud, WeatherKit, connected providers, web links, and support can still use a network. Each connection should be documented separately.
Is Apple Intelligence always processed only on my iPhone?
No blanket answer applies to every system feature. Apple says many requests run on-device and more complex requests may use Private Cloud Compute. An app using the on-device Foundation Models framework should describe its own feature path precisely.
Can an App Store privacy label prove that an AI app is private?
It is one useful signal, not complete proof. Compare the label with the product privacy page, permissions, included SDKs, offline behavior, optional services, export, and deletion controls.
How can I test whether a feature really works offline?
Complete any documented model setup, create non-sensitive test data, enable airplane mode, and attempt the core workflow from input to result and export. Record which steps still work and which explain that a connection is required.
Sources and further reading
Primary documentation is preferred. Product features and prices can change; verify details before deciding.