Apple's AI problem is not a Siri problem

The old Siri did not understand your sentences so much as match them. A request had to fall into a shape someone had anticipated in advance, or it failed — which is why a decade of patches never made it feel intelligent. Rebuilding it on a large language model, which is what Apple is reported to be doing for WWDC on June 8, removes that brittleness and installs a harder problem in its place: a system that can attempt anything can also be wrong about anything, fluently. How Apple manages that trade, not how the demo looks, is the real subject of this keynote.

I am fairly confident about the shape of what is coming, less so about the timing, and least so about whether it will work as shown. I will mark those differences as I go, because in this field the distance between a working feature and a demonstrated one is not a detail. It is the whole story.

Siri, rebuilt

The centrepiece is a reconstructed Siri: a dedicated app you can type into, a presence in the Dynamic Island, on-screen awareness that lets it act on what you are looking at, and 'personal context' drawn from your mail, messages, photos and calendar. The mechanism is what matters. The old assistant routed a fixed menu of intents; the new one is meant to reason over your screen and your data and then call the right action itself. Done well, that is the difference between a remote control and an assistant. Done badly, it is the old failures delivered in complete sentences.

The most telling detail is not a feature but a dependency. After its own models reportedly came up short, Apple is said to have agreed to run a custom version of Google's Gemini on its Private Cloud Compute servers to do the heavy reasoning behind the new Siri — the data kept inside Apple's perimeter, the intelligence rented from a competitor. Treat the specific partner as not yet confirmed and the arrangement itself as well-sourced; either way the asymmetry is the point. The company that sells privacy as the product is buying the capability it could not build in time.

iOS 27 is a consolidation year

Around Siri sits the annual sweep — iOS 27, iPadOS 27, macOS 27, watchOS 27, tvOS 27, visionOS 27 — and the notable thing is the absence of a headline. After last year's interface overhaul, this is widely briefed as a stability-and-battery release: performance, fixes, and refinement rather than reinvention, with incremental intelligence around the edges — better Genmoji and Image Playground output, Shortcuts that can call a model, reworked AirPods controls, wider live translation. That is sound engineering discipline. It also means that, in practice, the parts of the system you touch hourly will not close their distance from Android this year. The ambition is concentrated almost entirely in Siri, the one component with the worst delivery record.

The lever developers will watch

Quieter, and more consequential than any single feature, is the model Apple lets other apps call. Having begun exposing its on-device foundation model to developers, Apple is expected to widen that access: a more capable model, broader entitlements, perhaps a route to its server tier. The logic is sound — if developers can build useful, private AI on Apple's models without shipping their users' data to a third party, the platform compounds in a way a demo cannot. The constraint is equally plain. A framework is only as good as the model beneath it, and Apple's on-device model is small by construction. Developers will get integration and privacy; whether they also get capability worth choosing over a frontier API is the question that decides whether anyone builds on it.

Health, hardware, and the thinner rumours

Two further threads, and here the reporting is softer — treat them as plausible rather than settled. The first is a conversational health 'coach' that folds fitness, sleep and nutrition into personalised guidance, possibly paid; it is the one place Apple's private, on-device data could do something a general chatbot cannot. The second is hardware at the margins — a 'MacBook Ultra', the next turn of Apple silicon — alongside the conspicuous absence of the dedicated AI home device long rumoured and, it seems, delayed again. Hardware is the part of this story no one doubts Apple can execute, which is exactly why it is not the interesting part.

The promise already on the record

Scepticism about a Siri relaunch is not contrarianism; it is provenance. Apple has stood on this stage before and promised a more personal, context-aware Siri that could act across your apps — demonstrated it, advertised it, and then let the most ambitious pieces slip out of the release and out of the advertising. When a company shows a capability, sets expectations against it, and then cannot ship it, the burden of proof on the second attempt is not ordinary. Apple set that burden itself, and it set it high.

The frontier labs are judged on what they release. This year, Apple is asking to be judged on what it intends to.

The doubt is structural, not seasonal

The question underneath the keynote is whether Apple's way of building software is compatible with how this technology is built. Three mechanisms, not moods, drive the doubt.

The first is cadence. Apple ships intelligence once a year, in a single release gated by a marketing calendar; the frontier labs ship a meaningful change every few weeks and learn from use in between. A flawless WWDC still fixes Apple's response time at twelve months in a contest whose leaders move monthly. That is an asymmetry of clock speed, and it does not yield to effort.

The second is the on-device doctrine. Apple's privacy commitment is real and, to my mind, admirable, but by construction it favours small models on the phone, a Private Cloud Compute tier behind them, and a rented model behind that for anything genuinely hard. Capability at the frontier still scales with size and compute. Being the most private assistant and the most capable one are, for now, in tension — and Apple has chosen private. That may prove the right long-run bet. It is still a constraint, not a free choice.

The third is dependence, which the Gemini reporting makes concrete. For two years Apple's headline AI has run on another company's model. There is no shame in partnering, and routing it through Private Cloud Compute protects users. But a company renting the intelligence at the centre of its flagship is, by definition, not competing at the frontier; it is paying a rival for access to it. Accountability for the result, it is worth noting, still lands on Apple — which is the part of the arrangement Apple cannot outsource.

The case for Apple, stated fairly

It would be lazy to write Apple off, and the honest counter-case is strong. Distribution is the advantage no lab has: Apple can place a competent assistant in front of more than a billion people with no install and no sign-up. Its silicon team is the best in the world at capability per watt. And the research has drifted Apple's way — small models are far better than they were eighteen months ago.

The strongest version of the argument is one I think is correct: most people do not need a frontier model. They need a dependable one wired into their actual life — calendar, threads, photos, habits — that does ordinary things without drama. An assistant that privately and reliably knows your world could be more useful than a cleverer one that knows none of it, and Apple is, in principle, the company best placed to build it. Every clause of that rests on one word: if. If Siri ships. If it is fast. If you can reach for it without bracing for disappointment.

What to watch on June 8

Set aside the edited reel and watch for three things, because they are what separate a working system from a demonstrated one. Whether Siri is shown completing real, multi-step tasks live rather than in cuts. Whether Apple commits to a ship date inside this calendar year rather than the soft 'later this year' that has become its tell. And whether the demonstrations run on a real person's messy data rather than a stage account that never misbehaves.

Clear that bar and the comeback is earned, and this column ages badly — a fine outcome for anyone who owns an iPhone. Fall short, and the keynote will still have answered the only question worth asking, just not the one on the slide. Not whether Siri is better, but this: when the intelligence at the centre of the iPhone is rented, and designed on a clock it cannot match, who is Apple actually still in a position to compete with — and for how long?