Apple Intelligence had one of the messiest launches in recent Apple memory. Announced at WWDC in June 2024, it didn’t reach a single iPhone until October, and the headline feature, a Siri with real personal context, slipped far past that. Plenty of people installed iOS 18.1, poked at the new buttons, and quietly forgot about it.
That would be a mistake. Two years of updates later, it has become the sort of software you notice most when it’s gone: a rewritten email you didn’t have to rewrite, a 40-message thread collapsed into three lines, a photo with a stranger’s shoulder removed cleanly. None of it is flashy. Most of it works.
What Apple Intelligence actually is
It isn’t an app. There’s no icon called Apple Intelligence on your home screen and no chat window waiting for prompts. It’s a set of models woven into iOS, iPadOS and macOS that surface inside Mail, Notes, Messages, Photos and Safari, plus any third-party app that connects to Apple’s frameworks.
Two tiers do the work. A roughly 3-billion-parameter model runs on the device itself, handling small jobs: proofreading a paragraph, summarising a long thread, turning a cluttered notification into a readable sentence. Harder requests get routed to bigger models on Apple silicon servers, in a system the company calls Private Cloud Compute. For scale, that on-device model is a fraction of the size of the open-weight models that quietly power half the AI you already touch, which is why Apple leans so hard on routing rather than raw capability.
There’s also a ChatGPT handoff. When Siri or Writing Tools decides it’s out of its depth, you get a permission prompt before anything leaves the device. Decline and the request simply stops. Since 2025 you can point that handoff at other providers instead of OpenAI, if you’d rather.
The device list is narrower than the ads suggest
Apple Intelligence needs 8GB of unified memory. That means iPhone 15 Pro and Pro Max, every iPhone 16 and later, iPads with M-series chips, and Macs from the M1 onward. A standard iPhone 15 bought in 2023 gets none of it. Identical shape, similar price, different silicon. If you’re weighing an upgrade purely for these features, check the model number before anything else.
The features that earn their place
Writing Tools is the workhorse. Select text anywhere, including in third-party apps, and you get options to proofread, rewrite, shorten, or shift the tone between friendly, professional and concise. The summarise option is the one people end up using daily, because it turns a five-paragraph reply into a sentence you can act on.
Notification summaries are the most divisive. Apple temporarily pulled them for news and entertainment apps in early 2025 after the BBC flagged a summary that contradicted the story it was describing, then brought them back with clearer labelling and an option to switch them off per app. For message threads and calendar churn they’re genuinely useful. For anything involving nuance, they flatten it.
A handful of others have stuck:
- Clean Up in Photos, which erases objects from a shot with far less smearing than the older retouch tools.
- Call recording and transcription in Notes, with an automatic summary at the top of the transcript.
- Visual Intelligence, triggered from the Camera Control button on newer iPhones, which reads a poster, plant or restaurant front and offers an action.
- Genmoji and Image Playground, which are silly, occasionally delightful, and best treated as a messaging toy.
If that list feels familiar, it’s because these small assistive features are exactly what most people’s real AI usage looks like, even when they’d never describe it that way. A lot of it hides inside ordinary apps doing unremarkable things slightly faster.
Siri is still the part catching up
The original pitch promised a Siri that knew your calendar, your messages and your screen, so you could say “when is Mum’s flight landing” and get an answer pulled from an email. That version has been repeatedly delayed, and what shipped first was closer to a cosmetic refresh: a glowing edge around the screen and better handling of verbal stumbles.
The groundwork, though, is real. Apple’s App Intents framework lets developers expose their app’s contents and actions to the system, which is how Siri and Spotlight can eventually reach inside things they don’t own. When that lands properly, the assistant stops being a voice search box and starts being a way to operate your phone without tapping. Until then, expect inconsistent results and keep your expectations calibrated.
Privacy is the actual differentiator
Apple’s argument is that most processing never leaves the phone, and that the server-side part is auditable. Private Cloud Compute runs on Apple silicon in data centres, keeps no copies of your requests, and publishes software images that independent researchers can inspect. Whether or not you take the marketing at face value, the architecture is meaningfully different from shipping every query to a third party by default.
The practical effect: the permission prompt is a real gate, not a formality. If you decline the ChatGPT handoff, the feature degrades rather than silently proceeding.
Where it falls short
Writing Tools has a house style, and it is bland. Rewrite a punchy paragraph and you’ll often get something smoother, longer and less interesting, full of the connective tissue that makes corporate writing indistinguishable from itself. It’s a useful first pass and a poor final one.
Summaries hallucinate occasionally, particularly with anything involving numbers, negatives or quoted speech. Image Playground’s output is clean but sterile, with the flat sheen of stock illustration. And there’s no persistent memory across apps: it won’t remember that you prefer short emails, or that a colleague’s name is spelled unusually.
Those gaps matter less than they sound, because the tool is designed for narrow tasks rather than open-ended conversation. The honest assessment after two years of use is that it’s a competent utility with a handful of standout moments, not a assistant that reorganises how you work.
Practical ways to get more out of it
- Select text in any app and use the floating Writing Tools menu instead of opening the share sheet. It works in Slack, Gmail and most text fields.
- Turn on notification summaries only for messaging and calendar apps, and switch them off for news.
- Record interviews and meetings in Notes rather than a third-party recorder, then edit the auto-summary before sharing it.
- Use Clean Up on the edges of a photo, not the centre, where it has less context to invent.
- Type a Genmoji prompt with a specific object and setting rather than an abstract mood. The results improve noticeably.
How it stacks up against everything else
Google and Samsung have taken a louder approach, putting Gemini and Galaxy AI front and centre with dedicated apps and cross-app agents. Apple’s version is quieter and more conservative, which cuts both ways: fewer wow moments, fewer embarrassing misfires.
The competitive pressure is real, though. Open-weight model families now run locally on hardware Apple doesn’t control, and the gap between a phone-sized model and a server-sized one keeps narrowing. Apple’s bet is that tight integration and privacy beat raw benchmark scores. That bet looks better in year three than it did at launch, but it’s still a bet.
What developers build next will decide this
The most interesting thing about Apple Intelligence isn’t any single feature. It’s the Foundation Models framework, which lets third-party apps call the on-device model directly, at no cost, without sending data anywhere. A recipe app that reformats instructions for the ingredients you have. A note-taking tool that summarises a meeting as it happens. A journal that asks better questions because it can read what you wrote last week.
Apple has done the unglamorous work: the silicon, the routing, the privacy plumbing, the frameworks. Whether any of it feels essential in two more years depends on what thousands of developers do with that access, not on how good the next Siri demo looks on a keynote stage.