Apple brings generative AI to iPhone, opens Siri to ChatGPT

On June 10, 2024, Apple introduced Apple Intelligence for iOS 18, iPadOS 18 and macOS Sequoia, together with optional ChatGPT integration in Siri and Writing Tools. Apple’s release describes a hybrid system: some tasks run on the device, others use Private Cloud Compute, and ChatGPT requires permission before receiving a request.

The announcement matters for a simple reason: Apple is not presenting AI as a standalone app or a chatbot users have to seek out. It wants AI to work within the features millions of people already use to write, search for photos, manage notifications and talk to Siri.

AI designed around personal data

Apple Intelligence combines models that run directly on the device with requests sent to servers when a task requires more computing power. The company calls this infrastructure Private Cloud Compute: servers running on Apple chips that, it says, will process requests without retaining user data and use software that independent researchers will be able to inspect.

That distinction matters. AI assistants typically need access to email, calendars, documents and messages to be genuinely useful, but that same information is especially sensitive. Apple is trying to turn privacy into a competitive advantage over services that centralize such data in the cloud.

The system will initially be available in U.S. English and arrive in beta this fall. It will be compatible with the iPhone 15 Pro and 15 Pro Max, as well as iPads and Macs with an M1 chip or later. That leaves out a large share of current iPhones: owning a recent phone will not be enough if it lacks the required processor.

Text, images and a context-aware Siri

The most immediate features are Writing Tools, which can rewrite, proofread and summarize text in apps such as Mail, Notes, Pages and many third-party applications. Users will be able to request a change in tone, turn a draft into more concise copy or generate a summary of a message thread.

Apple also unveiled Image Playground, a feature for generating illustrations in styles such as animation, illustration and sketch. It is joined by Genmoji, which lets users create emojis from a description. In Photos, AI will be able to remove background elements and create memory videos from a written prompt.

The most ambitious overhaul affects Siri. The assistant will be able to maintain the thread of a conversation, understand references to items displayed on screen and use information spread across apps. Apple showed, for example, requests to find flight details sent by email and then check the time of a reservation.

This is more than an upgrade for answering questions: the goal is for Siri to take action across apps. The company said this capability will roll out gradually over the coming year, an important caveat because coordinating actions involving personal data is considerably more complex than summarizing text.

ChatGPT joins as a fallback, not the primary assistant

When Apple Intelligence or Siri cannot handle a request on their own, they will be able to offer to send it to ChatGPT, initially powered by GPT-4o. Users will see a confirmation prompt before sharing the question or documents with OpenAI.

ChatGPT will also be integrated with Writing Tools to draft or transform text. Apple said it can be used without creating an account and that IP addresses will be hidden; people who already have a paid ChatGPT subscription will be able to connect their account to access the features included with their plan.

The partnership is significant for both companies. For Apple, it avoids having to build its own response to every open-ended or creative query from day one. For OpenAI, it puts ChatGPT inside Apple’s operating systems, although the experience will still be presented as a device feature rather than as a rival app to Siri.

The real test will be delivering beyond the stage

Apple is arriving later than Microsoft, Google and Samsung to the race to bring generative AI to computers and phones. Its strategy is not to compete for the most visible chatbot, but to turn AI into a layer that understands the user’s context and works with fewer steps.

The approach could prove more practical than a standalone assistant if Siri truly understands what is on screen and can connect information across apps. But it also requires Apple to show that these automations are reliable, that cloud requests uphold the promised safeguards and that the features arrive with enough quality for everyday use.

The beta launch, limited to U.S. English, will make it possible to assess that balance first. For everyone else, the big question is no longer whether Apple will bring generative AI to its devices, but when it will reach their languages and how much of the promise shown today will be available from the initial launch.

A hybrid assistant needs a data map

The question “does it run on the phone or in the cloud?” can have several answers within one interaction. The system may classify a request locally, consult personal data, send part to an Apple server and ask permission to forward another part to ChatGPT. A privacy assessment must follow every hop: content sent, provider, purpose, retention, identifiers and available controls.

On-device compute also depends on hardware. A model compatible with a particular processor does not mean identical features across an entire product range. The record should separate operating system, device, language, region and availability status. In an announcement, “will come” describes a plan; “is available” describes a testable product. Combining them makes a future demonstration look like a present capability.

How to test a Siri action

For one task, first write down the data it needs and the result it should produce. Then repeat it offline, with restricted permissions and with ambiguous information. Observe whether the system explains which service it will use, asks before sharing a file and leaves a comprehensible record. A polished answer does not justify sending more context than necessary.

Actions across applications carry a different risk from text generation. Finding an email is reversible; sending a message, changing an appointment or sharing a photograph may not be. Sensitive operations require a preview, visible recipient and confirmation. Useful personalisation should rely on narrow permissions, not permanent access to all personal context.

The transferable skill is to turn a promise of “personal AI” into a diagram of routes and permissions. Mark what happens locally, what leaves the device, who receives it and which action can be executed. That map compares assistants even when their model, provider or brand changes.

Permission must arrive at the right moment

A generic authorisation during setup is forgotten; confirmation immediately before sending a photograph or document exposes the consequence. A useful notice identifies the service, content and action, and allows the user to continue without the outside provider. There should then be a way to review enabled integrations and revoke access without losing the rest of the assistant.

A feature record should preserve device, version, language, region, data route and authorised action. Retest when an update changes any of them. “Apple Intelligence” is a family of paths, not one uniform property that can be approved once.

Consent must follow the route rather than be assumed.