Shipping AI in iOS Apps Without a Rejection: Guideline 5.1.2(i) Explained
18 Jun 2026 · 5 min read · Erik van Kammen
Quick answer: Apple's Guideline 5.1.2(i), added on November 13, 2025, says you must clearly disclose where personal data will be shared with third parties, including third-party AI, and obtain explicit permission before doing so. In 2026 this collides with the new Foundation Models framework, which can route prompts to Apple's on-device model, Apple's Private Cloud Compute, or a third-party cloud such as Claude or Gemini. If user data leaves the device to a third-party AI, you need named disclosure and explicit consent, and your App Privacy label must match. Here is how to check your build before you submit.
Adding an AI feature to an iOS app has never been easier, and that is exactly the problem at review time. A single API call can send a user's text, photos, or files to a model running in someone else's cloud, and Apple now has a specific guideline about it.
What does Guideline 5.1.2(i) actually require?
The rule states that you must clearly disclose where personal data will be shared with third parties, including with third-party AI, and obtain explicit permission before doing so.
In practice, as reported when the rule landed, "explicit permission" means a real consent step, not a line buried in your terms of service or privacy policy. A generic statement like "we share data with service providers" is not enough. The expectation is that you name the third party (or the category of third-party AI) and let the user agree before their data is sent.
This sits on top of your existing obligations: the App Privacy label on your product page and your privacy manifest must reflect the data you actually collect and share.
Why the 2026 Foundation Models framework makes this easy to get wrong
At WWDC 2026, Apple expanded the Foundation Models framework so a single Swift API can call the on-device Apple Intelligence model, Apple's Private Cloud Compute server models, or a third-party provider such as Claude or Gemini through a new language-model protocol. That is powerful, but it means the destination of a prompt can change with a configuration value.
The compliance line is about where the data goes:
- On-device only: personal data does not leave the device, so 5.1.2(i) third-party sharing is not triggered by the model call itself.
- Apple's Private Cloud Compute: this is Apple's own infrastructure, not a third party.
- A third-party cloud (Claude, Gemini, or any other provider): personal data is shared with a third party, so disclosure and explicit consent apply.
The risk is shipping a routing default, or an SDK, that quietly sends user content to a third-party model while your consent flow and App Privacy label still describe an on-device experience.
When the disclosure rule applies
| Data flow in your app | Does 5.1.2(i) apply? | What you must do |
|---|---|---|
| Model runs fully on device | Not for the model call itself | Keep the App Privacy label accurate |
| Routed to Apple Private Cloud Compute | No third-party sharing | Disclose data handling as normal |
| Sent to a third-party AI cloud (Claude, Gemini, others) | Yes | Name the provider, get explicit consent, reflect it in the App Privacy label |
| Third-party AI SDK bundled by a dependency | Yes, if it sends personal data | Inventory the SDK, confirm the data path, disclose and consent |
How to audit your AI data flows before submitting (checklist)
- List every place the app sends user-provided content (text, images, audio, files) to a model.
- For each, record the destination: on device, Apple Private Cloud Compute, or a named third-party cloud.
- For any third-party destination, add an explicit consent step that names the provider before data is sent.
- Update the App Privacy label and privacy manifest so the declared data collection and sharing match the real flows.
- Check transitive dependencies: an analytics or feature SDK may call a third-party AI on your behalf.
- Keep a short note in App Review describing your AI usage and where data goes.
Where a pre-submission check helps
This is a binary-level question, which is where a pre-submission scan is useful. A scan can inventory the SDKs and network destinations present in the compiled app, including AI providers pulled in by a dependency, and flag where your declared App Privacy data does not match what the binary can actually do. That is the gap that turns into a 5.1.2(i) rejection. For the privacy-label mechanics, see our guide on privacy manifest validation, and for the broader AI rules, Apple Foundation Models and AI compliance.
Frequently asked questions
Does an on-device AI model trigger Guideline 5.1.2(i)?
The model call itself does not, because personal data stays on the device and is not shared with a third party. You still need an accurate App Privacy label for any other data the app collects.
Is naming "a third-party AI provider" enough, or do I name the company?
The rule requires clear disclosure and explicit permission before sharing. Reporting on the rule indicates you should name the provider or the specific category of third-party AI, and a generic "service providers" statement is not sufficient.
Can I put the AI disclosure only in my privacy policy?
No. The expectation is explicit permission, which means a consent step the user actively agrees to, not a clause buried in the terms of service or privacy policy.
Does using Apple's Private Cloud Compute count as third-party sharing?
No. Private Cloud Compute is Apple's own infrastructure. Sending data to a non-Apple cloud such as Claude or Gemini is third-party sharing and triggers the disclosure and consent requirement.