Overview
Apple Intelligence is the set of generative features Apple has been building into iOS, iPadOS and macOS since WWDC in June 2024, shipped in stages from iOS 18.1 onward. It covers writing tools, notification and mail summaries, Genmoji, Image Playground, and, since a beta that began on September 14, 2026, a rebuilt Siri that reads what is on your screen and performs actions inside apps. It runs only on iPhone 15 Pro and iPhone 16 or later, wants roughly 7GB of free storage, and is off until you switch it on.
Two things set it apart from everything else rated here. The first is architectural. Most requests are answered by models running on the phone, and requests that need more go to Private Cloud Compute, Apple’s own server tier, which Apple says processes them without writing user data to persistent storage and without Apple being able to read them. Apple publishes code measurements for every production PCC image to an append-only transparency log, has released source for parts of the stack, shipped a Virtual Research Environment that lets outsiders boot and prod a real release on their own Mac, and pays between $50,000 and $1,000,000 for PCC vulnerabilities. Nothing else on this Index has invited that kind of inspection.
The second is that the marketing outran the product by enough to cost real money. Apple delayed the promised personalised Siri in March 2025 and pulled the ads. In May 2026 it agreed to pay $250 million to settle a class action alleging it misrepresented what Apple Intelligence would do around the iPhone 16 launch, without admitting wrongdoing. Claims opened on September 21, 2026, for buyers of an iPhone 15 or 16 between June 10, 2024 and March 29, 2025, at up to $95 per device.
What It Knows About You
The honest answer depends on which feature you turn on. Writing tools, summaries and Genmoji are largely local, and Apple says it does not train its foundation models on your device data. The 2026 Siri is a different proposition: it builds a personal-context index across messages, mail and photos, reads on-screen content when you ask about it, and invokes actions that developers expose through App Intents, including actions in apps you never opened. That is close to full read access to a phone, handed to a system service, in exchange for being able to ask in plain language.
Where a request does leave the device it goes to PCC, which Apple describes as stateless. Since June 2026 some of that capacity sits in Google Cloud data centres on NVIDIA hardware rather than only on Apple silicon in Apple’s own buildings, under what Apple says are the same guarantees. Separately, Apple agreed in January 2026 to license Google Gemini models, reported at roughly $1 billion a year, to underpin the next generation of Apple Foundation Models. Apple’s position is that the model weights run inside PCC and that Google is contractually barred from training on the queries. That is Apple’s account of a private contract, not something an outsider has verified.
The seam is the ChatGPT hand-off. If Siri routes a question to OpenAI and you are not signed in to an OpenAI account, Apple says the request carries nothing tied to your Apple Account and OpenAI may not store it or train on it. Sign in, and OpenAI’s own terms govern instead. The protection is real, and it is one tap deep.
The Real Risks
Confident fabrication, well documented. The notification summaries in iOS 18 rewrote BBC alerts into events that had not happened: that Luigi Mangione had shot himself, that Luke Littler had won the world darts final hours before it began, that Rafael Nadal had come out as gay. After the BBC complained, Apple disabled summaries for news and entertainment apps in January 2025. They came back in iOS 26 with a red-text warning that summarisation may change the meaning of the original headline. The warning is an improvement and also an admission.
Bias in image generation, measured from outside. Machine learning researcher Jochem Gietema tested Image Playground in February 2025 and found it produced lighter skin tones for prompts such as affluent, rich and successful, and darker tones for poor, impoverished and destitute, with comparable patterns across prompts about sports, jobs and clothing. Apple restricts Image Playground to cartoon styles partly to contain this. We found no published Apple bias evaluation to weigh against the outside testing.
Consent that did not stay given. People who had switched Apple Intelligence off found it back on after updating to iOS 18.3.1, and again after 18.3.2. The mechanism was a post-update welcome screen offering Continue with no Set Up Later, which functions as a dark pattern whether or not it started as a bug.
Unresolved claims about the training corpus. Authors sued Apple in September 2025, and in a separate complaint on October 9, 2025, alleging that Applebot ingested pirated books from shadow libraries. Those are allegations; no court has ruled on them.
The old Siri’s record belongs in the ledger even though it predates these features. Apple agreed to pay $95 million to settle claims that Siri recorded private conversations after unintended activations between September 2014 and December 2024, denying any wrongdoing. Payments began in January 2026, capped at $20 per device.
Verifiability has a ceiling, too. Researchers at TU Darmstadt analysing PCC reported that the published binaries are not reproducibly built and carry no symbols, so an outsider cannot conclusively confirm that the published source is what runs in production. That is a gap in the strongest claim Apple makes here rather than a refutation of it.
Alternatives
- Leave it off. Apple Intelligence is opt-in, and switching it off removes the on-device models. Re-check the setting after each major iOS update.
- Take some features and not others. Screen Time has separate controls for Writing Tools, Image Creation and Intelligence Extensions, under Content and Privacy Restrictions.
- Turn off notification summaries for news apps specifically. The accuracy failures were concentrated there, and the setting is per-app.
- Decline the ChatGPT extension, or use it without signing in to an OpenAI account. Signing in is the step that moves your request under OpenAI’s terms.
- Write the summary yourself. A paragraph you wrote is one you can vouch for, and no draft of your own has ever invented a headline about a real person.
Our Verdict
B, the second one on this Index. Microsoft Recall earned its B by being narrow and local; this earns the same grade on architecture, while doing far more. Privacy scores 3, the lowest we have given anything, because Apple kept the default on the device, built a server tier that stores nothing, and then did the uncomfortable part: publishing the transparency log, releasing source for security-critical components, and paying up to $1 million to whoever breaks it. Job threat scores 2, since writing tools and Genmoji displace almost no paid work.
The grade is not higher because of autonomy and accuracy. The 2026 Siri reads your screen, indexes your mail and photos, and chains actions across apps on one spoken request, which is a genuine transfer of judgement and why autonomy sits at 6. The summaries invented events involving named people, and the remedy shipped was a disclaimer. Switching the whole thing off did not reliably stay off across two 2025 updates.
What would raise the grade: reproducible builds of the PCC images, so the source claim can be checked rather than trusted; a published bias evaluation of Image Playground; and an opt-out that survives a software update. What would lower it: an adverse ruling in the authors’ copyright cases, any evidence that PCC retained request data, or a repeat of the January 2025 summary failures now that news summaries are switched back on. As of October 2026 the rebuilt Siri is still in beta, so the autonomy score is rating a product that is still arriving.