Apple has not yet disclosed which of the new accessibility capabilities stay on device and which may route to Private Cloud Compute, leaving the most sensitive user category to wait for the privacy documentation that ships with the release.
On May 19, 2026, Apple previewed a suite of accessibility updates that lean on Apple Intelligence across VoiceOver, Magnifier, Voice Control, and Accessibility Reader, plus a Vision Pro feature for controlling compatible power wheelchairs with eye gaze and on-device generated subtitles for uncaptioned video across the Apple ecosystem. Most coverage treated this as a Global Accessibility Awareness Day feature parade. The more honest read is that the suite is a stress test, not a celebration, of the "on-device by design" claim Apple has been building toward for two years.
That is because the people who depend on these features are exactly the ones who cannot tolerate sensitive context leaving the device. A blind user asking VoiceOver's new Image Explorer to describe a scanned medical bill, a hard-of-hearing user watching live captions track a noisy restaurant conversation, a Vision Pro wearer steering a wheelchair through a crowded lobby: each of these is a context where the cost of a cloud round-trip is not just latency, it is the wrong default for what the user has asked the system to see. Accessibility is the cleanest available check on whether Apple's on-device promise is a principle or a marketing line, and Apple has not yet published the architectural map that would let the reader verify it.
The features themselves are substantive. VoiceOver gains what Apple calls VoiceOver Image Explorer, a detailed-description mode that walks through photos, scanned bills, and personal records, with natural-language navigation for users who are blind or have low vision. The Magnifier app picks up new ways to describe surroundings and read text in view. Live Recognition and Voice Control get updates under the same Apple Intelligence umbrella. Accessibility Reader, Apple's text-and-speech reading view, gets refinements. On-device generated subtitles arrive for uncaptioned video content across iPhone, iPad, Mac, and Apple TV. Vision Pro learns to control "compatible power wheelchairs" through eye gaze, a feature The Register called the first consumer-facing example of eye-gaze wheelchair control from a major platform vendor.
The Verge's coverage was more measured, walking through each capability without inflating it. That is the right instinct. The work is real: Apple's accessibility team has shipped a steady cadence of features for years, and the new round continues that pattern. Sarah Herrlinger, Apple's senior director of accessibility policy, has been the public face of that work, and Tim Cook used the announcement to frame the suite as proof of a "privacy by design" stance across the company's AI roadmap.
That framing is the part that needs scrutiny. Cook's quote, and Apple's, treats on-device processing as a settled architectural fact. It is not yet, for this suite, a disclosed one. Apple has historically split Apple Intelligence workloads between on-device inference and a service it calls Private Cloud Compute, which routes harder requests to Apple silicon servers while preserving what the company describes as a privacy audit trail. For features that involve describing personal photos, transcribing spoken context, or steering a wheelchair in real time, the difference between those two paths is not subtle: one keeps the user's data on the device, the other asks the reader to trust a server-side architecture that Apple itself has argued is uniquely auditable but that no independent party has been given the ability to inspect at the per-feature level.
As of the May announcement, Apple has not published a per-feature breakdown of which accessibility capabilities run on-device and which may route to Private Cloud Compute. The newsroom post describes "Apple Intelligence" as the engine and "later this year" as the ship date, but the architectural map the privacy promise depends on is forward-looking. The Register and The Verge both flagged the on-device framing in their write-ups without resolving it. Apple's own May 19 newsroom post does not enumerate the split per capability.
This is the gap that should concern the reader more than any single feature does. Accessibility is the category that exposes the cost of an unverified architectural claim, because the users involved are the ones for whom the cost of being wrong is highest. A privacy promise that holds for a Magic Eraser-style photo edit but routes a VoiceOver description of a medical document off-device is a promise with a hole in the middle. Apple has earned the benefit of the doubt on accessibility execution across years of shipped features. It has not yet earned the conclusion that the on-device promise for this suite is fully kept, because the company has not given the reader the map that would let them check.
What to watch: when Apple ships these features later in 2026, the privacy documentation that accompanies the release will be the first real test. If Apple publishes a per-feature split, on-device versus Private Cloud Compute, the architectural claim becomes verifiable. If it does not, the suite ships as advertised and the principle test fails quietly. For a category where the users most at stake are the ones with the least leverage to demand a follow-up, that distinction is the story, not the feature list.