原文 · 未翻译
AI
Report
Tech
Apple’s AI pitch will live or die by its privacy promise
Apple doesn’t have the latest or greatest AI, but it might have the most private.
Apple doesn’t have the latest or greatest AI, but it might have the most private.
As expected, yesterday’s WWDC keynote was mostly about AI. And also as expected, Apple tried to turn its late arrival into its sales pitch: It didn’t rush into AI because it was taking its time to do things right. In this case, “right” means “with more privacy than anyone else.” It’s a good pitch — the question will be how well it holds up.
The new Apple Intelligence features and the updated Siri AI have been designed to work across iPhone, iPad, Mac, Apple Watch, and Vision Pro. There’s a dedicated Siri AI app, with a ChatGPT-esque chatbot experience, new AI-powered camera and photo editing features, and the beginnings of an agentic experience that will let Siri AI interact with other apps and software on your iPhone, iPad or Mac.
Whatever device you access the new AI from, Apple says the processing will be roughly the same: Queries will be handled on-device where possible and in its secure Private Cloud Compute system when not. Apple says your data won’t be stored, will only be used to execute your request, and won’t be accessible to Apple or anyone else. Conversation logs in the new Siri AI app will be kept, but only on-device and in your end-to-end encrypted iCloud account.
This architecture itself isn’t new — Private Cloud Compute was announced alongside the initial Apple Intelligence launch in 2024. But two years on, two things have changed. First, Apple is undeniably behind almost every competitor on AI, even after yesterday’s announcements. That makes its privacy pitch more important than ever as a way to differentiate itself from the competition. It’s the second change that makes that trickier: In part because it’s so behind, Apple is now working with Google and Nvidia to run its AI operation.
Instead of being designed entirely in-house, Apple’s new cloud AI models are based on Google Gemini. Meanwhile Private Cloud Compute has expanded beyond Apple’s own data centers, running on Google Cloud systems using Nvidia GPUs, Intel CPUs, and Google Titan chips.
That’s a marked change. When Private Cloud Compute was first announced, Apple emphasized that it was built specifically to run on Apple silicon, with a hardened supply chain including extensive security scans and validation checks on each server before it joins the rack. Apple can’t control Google, Intel, and Nvidia’s supply chains though. Instead, Apple now maintains a “cryptographically verifiable, append-only ledger” of all Google Cloud hardware used for Private Cloud Compute and “retains complete control” of the software. Apple claims the resulting system has the same “extraordinary security and privacy properties” as before, though skeptics might point out that the longer supply chain introduces vulnerabilities that simply didn’t exist before.
Still, Apple can credibly claim that its approach to AI places a priority on privacy lacking from the AI giants. And it wasn’t shy about making that point. “Some appear to be racing forward, seemingly pursuing AI for the sake of AI, without clear regard for the people, all of us, that it’s ultimately meant to serve,” SVP of software engineering Craig Federighi said during the keynote. He went on to explain that Apple Intelligence has instead been designed “with privacy in mind at every step.”
In truth, Apple’s privacy solutions aren’t entirely unique, though they do remain the industry’s most comprehensive. Last year Google announced Private AI Compute, which at the time we called “virtually identical” to Private Cloud Compute (right down to the name), running on “one seamless Google stack” powered by its in-house TPUs. The difference is in application: While Apple uses its private cloud computing for every AI query that can’t be run on-device, Google has been less specific about when it uses Private AI Compute — and when it doesn’t. The company says it’s used for Magic Cue and the AI-powered Recorder app on Pixel phones, but hasn’t said if it’s used for every Gemini query made from one of Google’s phones, or if there are any equivalent protections for Gemini use on other platforms.
In fact, Google says that Gemini collects reams of data by default: prompts you submit, files you share, and recordings of spoken conversations. Google also collects data on the content Gemini creates for you, the tasks it carries out, and information from and about the apps, browsers, and devices you use it with. By default, chat history with Gemini is stored for 18 months and then deleted, though you can reduce this to as little as 72 hours.
It’s a similar story with other major AI companies. OpenAI says it collects the prompts and content that users upload to ChatGPT, along with various types of location and device information. Chats are used as training data by default, though you can disable this. Anthropic collects roughly the same stuff when you use Claude, though notes that it deletes audio recordings of spoken prompts, while retaining the transcripts. Claude also defaults to using your data for model training and says it’s kept in a “de-identified” form for up to five years.
By contrast, the Apple Intelligence privacy policy — last updated in 2025, before the most recent announcements — says that Apple only collects “limited information” on Private Cloud Compute requests, such as the size and how long it takes to complete, but no information about the content of the request or its result. Apple says it doesn’t use private data or user interactions to train its foundation models. Of course, Apple has an advantage here: It doesn’t need to collect your data to train its models, because Google has already handled that using its users’ data. Apple’s partnership with Google may raise concerns about how it maintains user privacy, but it may also be precisely how it’s able to offer functional AI without giving up that privacy.
It’s the more limited default data collection that ultimately makes Apple’s privacy promise stand out. Even if the expanded Private Cloud Compute turns out not to be as secure as it used to be, Apple is still claiming to collect a lot less data than any of the AI alternatives. For many Apple users, that may be worth waiting a few months — or years — for Apple Intelligence to finally arrive.
Dominic Preston
AI
Anthropic
Apple
Apple Event
Google
OpenAI
Report
Tech
More in: WWDC 2026: All the news from Apple’s developers conference