How to Set Up Apple Intelligence and Keep More of It On-Device

Photo by Fujiphilm on Unsplash

Apple Intelligence ships with a quiet default most users never change: requests that exceed the on-device model's capacity are silently escalated to Apple's Private Cloud Compute, and then optionally to ChatGPT. Learning how to set up Apple Intelligence on iOS 26 properly means understanding that escalation path and deciding, per feature, where you want your data to actually live. This guide walks through enabling the system, controlling the cloud handoffs, and fixing the one failure mode that traps most new users.

What Actually Happens When You Tap "Turn On Apple Intelligence"

Apple Intelligence is not a single model. According to Apple's Security Engineering team, requests run through a three-tier hierarchy: an on-device model handles what it can, more complex prompts are encrypted and routed to Private Cloud Compute (PCC) servers built on Apple silicon, and a third tier hands off to ChatGPT only when you give explicit consent. The transition between tiers is automatic and largely invisible in the UI.

The local model is small by design. SolidAI Tech reports that the on-device foundation model is roughly 3 billion parameters and downloads as a file of about 4 GB, with Apple recommending you keep the device on Wi-Fi and power during the initial pull. Anything that model cannot resolve gets escalated.

That escalation is the part worth slowing down for. PCC nodes run a hardened subset of iOS, intentionally stripped of remote shells and introspection tools, and traffic to them is end-to-end encrypted to the nodes' public keys so that load balancers outside the trust boundary cannot decrypt it, Apple's PCC documentation states. The system is also stateless: Apple says user request data is not persisted after a task completes, and access is cryptographically blocked rather than restricted by policy.

Confirm Your Device Can Actually Run It

This is the step that catches the largest share of failed setups. The Neural Engine gates everything, and Apple Headlines notes the minimum requirement is an iPhone 15 or newer, with Intel Macs now restricted to cloud-only features and Apple Silicon required for any on-device inference. An iPhone 14 Pro will not get Apple Intelligence regardless of the iOS version installed.

Before tapping anything, check four things:

• iOS version: iOS 26.1 is the floor, but SolidAI Tech recommends iOS 26.2, which resolved a battery-drain regression caused by background model inference processes.
• Free storage: at least 7 GB, per Apple Headlines, to fit the foundation model and its working files.
• Language match: your Siri language and device language must be set to the same supported language, or the model will refuse to load.
• Power state: Low Power Mode disables several AI features and pauses background model downloads.

If you skip the language check, you will hit a silent failure where the toggle exists, you can flip it, and nothing actually downloads. That is the single most common setup failure, and it produces no error message.

Turn It On Without Surrendering the Defaults

The enablement path is short. Open Settings, tap Apple Intelligence & Siri, then tap Turn On Apple Intelligence, as Apple's iPhone user guide describes. The on-device model starts downloading in the background, and Apple's setup guidance confirms download time varies, with Wi-Fi and power producing the fastest result.

A few specifics matter here. Apple does not charge for the feature on supported devices, Apple Headlines confirms, so any prompt asking for payment is not legitimate. The download will not begin if you are on cellular only, and switching off Apple Intelligence later removes the on-device models entirely, freeing the storage but forcing a full re-download if you change your mind, per Apple Support.

During first run, iOS does not ask you about ChatGPT. The ChatGPT extension is a separate opt-in buried under Extensions in the same settings pane, which is the right place to spend a minute before you start using Writing Tools or Siri.

Apple Intelligence On-Device vs Private Cloud Compute: Where Each Request Goes

There is no setting in iOS 26 that forces Apple Intelligence to stay strictly on-device. The OS decides per request based on model capability and prompt complexity, and the user is not asked. This is the most significant transparency gap in the product as shipped.

What you can control is the upstream behaviour. Requests handled locally never leave the Neural Engine. Requests escalated to PCC are encrypted directly to node public keys, and Apple's security post emphasises that the privacy gateway sitting between the device and the compute nodes cannot decrypt the payload. The cryptographic guarantee is stronger than what any mainstream competitor publishes, including Google's Gemini cloud path and Microsoft's Copilot routing.

The ceiling on this architecture is real. MacRumors reports, citing The Information, that Apple has struggled to run full Gemini-scale models on PCC because the infrastructure uses the same Apple silicon chips found in Mac computers, creating a hard limit on how large a model can serve a real-time request. That is why Apple is reportedly distilling a large Google Gemini model into a smaller version that can run locally, and why the company is positioning on-device AI as a headline focus for WWDC 2026 on June 8.

Lock Down the ChatGPT Handoff Before You Forget

ChatGPT is the only tier that leaves Apple's trust boundary entirely. The privacy story is still better than using the ChatGPT app directly: Apple strips your IP address and Apple Account from the request before it reaches OpenAI, so OpenAI cannot identify which Apple user submitted the query, SolidAI Tech notes. iDownloadBlog adds that OpenAI is also not permitted to store anonymous requests for model training, though if you sign in to a ChatGPT account the consumer privacy policy applies instead.

To configure the extension, go to Settings, tap Apple Intelligence & Siri, then under Extensions tap ChatGPT and tap Set Up, following the path Apple Support documents. Two choices follow. Tap Enable ChatGPT to use it without an account and preserve the anonymisation guarantee. Tap Use ChatGPT with an Account if you want chats saved to ChatGPT history, accepting that OpenAI's account-level data handling now governs those requests.

The more interesting toggle is Confirm ChatGPT Requests. With it on, Siri prompts you before every handoff. OpenAI's help documentation notes you can flip this off to bypass confirmation for text queries, while Siri will still always ask before sending any file or image. Leaving the prompt on is the right default for anyone treating ChatGPT as the third tier rather than the first.

The One Setting That Decides How Personal Siri Gets

Apple Intelligence's deeper Siri features are not yet shipping. The original WWDC 2024 announcement was followed by a prolonged delay to the more personal Siri capabilities, and MacRumors reports that WWDC 2026 is positioned to reframe that delay. SolidAI Tech notes that a full LLM-driven Siri overhaul with a redesigned interface is expected to be unveiled at WWDC 2026, with a possible iOS 26.4 developer beta release shortly after.

For now, the most consequential Siri setting under Apple Intelligence & Siri is the option that grants Siri access to on-screen context and personal data across apps. Granting it expands what the on-device model can do without escalation. Withholding it forces more requests up the tiers, which often means PCC. Privacy-conscious users instinctively turn it off and unintentionally push more of their data to the cloud as a result. That is the inverted incentive the current Settings pane does not explain.

When the Toggle Looks On But Nothing Works

The most common post-setup failure is a model that never finishes downloading. Three causes account for nearly all of them:

1. Siri language mismatch: setting Siri to one language and the device to another suspends model downloads, Apple notes in its setup guide.
2. Insufficient storage: under 7 GB free and the download silently pauses; the 4 GB model size from SolidAI Tech's measurement is the floor, not the working set.
3. Low Power Mode: enabled either manually or by automation, this halts background inference and download tasks.

If the toggle is on and Writing Tools still does not appear in the share sheet, force-quit Settings, plug into power, connect to Wi-Fi, and wait. The model download is not exposed as a progress bar in any user-facing screen.

What This Actually Costs You in Privacy Terms

Apple explicitly does not use user interactions to train its foundation models, SolidAI Tech confirms. That separates Apple Intelligence from most consumer AI assistants by default, and it applies whether a request is served on-device or by PCC. For ChatGPT-routed requests without an account login, the anonymisation layer means OpenAI cannot tie the query to you even if it logs the prompt.

The Neural Engine that makes this possible is not new. It has shipped in every iPhone since the A11 Bionic in 2017, per Apple Headlines, representing nine years of silicon investment before Apple shipped its first generative feature on top of it. The hardware was waiting for the software, not the reverse, which is part of why the on-device tier works at all on a phone-class power budget.

If you want to follow how the on-device versus cloud split evolves at WWDC 2026, the AI Infrastructure articles and Privacy & Security articles sections track the parts of this story that affect what your device sends and what it keeps. For a contrast in how a competitor handles the same routing problem, the breakdown of Google's AI Mode agents and personal intelligence is the closest parallel currently shipping. Then revisit your ChatGPT toggle whenever Apple ships an iOS 26.x point release, because the default behaviour around confirmations has changed once already.