New Siri Indexes Your Private Data. Now What?

The most revealing thing about the new Siri isn't that it finally works. It's why it didn't work before.

On a recent Vergecast episode, Nilay Patel and David Pierce worked through the post-WWDC Siri testing reports — including Wall Street Journal tech columnist Joanna Stern's hands-on impressions — and landed on a deceptively simple explanation. Apple, apparently, just built the on-device search index wrong. Not slightly wrong. Built-it-wrong-and-shipped-it-for-the-entire-run-of-modern-iOS wrong. (iOS launched in 2007, which depending on when you're reading this puts that stretch somewhere between 17 and 18 years — "nearly two decades" is in the right neighborhood, though the hyperbole is worth naming.) According to Patel, Apple's Mike Rockwell confirmed at a WWDC tech talk that they scrapped the whole thing and rebuilt from the ground up. New database architecture, full stop.

That's the actual story. Not a frontier AI breakthrough. A fixed database.

"I don't even think it's the AI that's doing it," Pierce said. "I think it's like functionally they rebuilt the index of content and now an AI can just go get it — and that is magic."

Patel pushed back, reasonably: the index and the model are two sides of the same coin. You need both. But his framing of where the AI lives matters a lot — and not just as a product question.

The privacy architecture nobody's explaining clearly

Here's the part that got glossed over in most of the reaction coverage, and it's the part I keep coming back to.

The new Siri isn't purely on-device. According to what Patel heard in briefings — and he was clear this is his read of unconfirmed details, not Apple's official documentation — the likely architecture works like this: you make a query, it goes up to Apple's Private Cloud Compute infrastructure, gets processed using what he speculated may involve a Gemini-derived model (Apple has been conspicuously vague about what exactly its Google partnership covers), and the cloud decides which pieces of your on-device index to pull. Your data stays local. The query logic doesn't.

Apple's official position is that Private Cloud Compute provides strong privacy guarantees — your data isn't stored, isn't used for training, and the computation happens in isolated hardware. Independent researchers have done some verification of these claims. That's meaningfully better than just sending your messages to a server farm.

But "meaningfully better" and "airtight" aren't the same thing. The Gemini angle specifically is worth flagging: Apple has not publicly confirmed what capabilities it licensed or how Google's technology is involved in processing. Patel's hypothesis about speech-to-text and natural language processing is plausible — Google is genuinely excellent at both — but it remains speculation from a well-sourced journalist, not confirmed reporting. If your threat model includes "I don't want Google-adjacent systems processing queries about my iMessages," that's a reasonable concern to hold until Apple gives a clearer accounting.

Practically speaking: you can go to Settings > Apple Intelligence & Siri > Improve Siri & Dictation and turn off the features that contribute data. You can also disable Siri entirely while keeping Spotlight search. What you can't currently do, as far as anyone has tested, is audit what the index has already catalogued on your device, or exclude specific apps from indexing while keeping others. That audit capability doesn't exist in the current betas. It should.

The lock-in math just changed

This is where the privacy and the product story converge in a way that should give people pause.

Patel and Pierce made the lock-in point clearly: iMessage has always been a retention mechanism, but a broken Siri undermined its full potential. Now that the index works, the stickiness compounds. If you've kept messages set to "never delete" — and a lot of people have, because storage got cheap and deletion feels like loss — you now have an AI that can synthesize years of your communication history. That's genuinely useful. It's also genuinely hard to walk away from.

"The lock-in feels stronger than ever if Siri is good," Patel said, "because good Siri is going to be very hard to walk away from and to start over from with one of these other things."

Switching to Android has gotten easier technically. But the cognitive switching cost — starting over with an assistant that has zero context on your life — is now much higher. Every AI service has been trying to build this kind of memory and context. Apple just activated it for a billion-plus users who never opted into an AI product, they just never deleted their messages.

The iOS 27 Siri redesign reportedly goes further, with deeper personal context and model choice built in. Which means this indexing investment is the foundation, not the ceiling.

For people who care about their data: the honest answer is that this is a reasonable trade-off if Apple's privacy architecture holds. The on-device index staying local is genuinely protective. The cloud compute piece is less clear. Watch what independent security researchers find when they get deeper access to the beta. Don't make permanent decisions about your message retention habits based on current enthusiasm.

What Mosseri actually said, and what it means for you

The Vergecast's social section covered three platform moves — Instagram algorithm controls, Bluesky's community schema extensions, YouTube DMs — and Patel wove them into a coherent theory about platforms trying to feel smaller. All true and worth reading. But the piece of it that actually matters for how you use your phone is buried in an Adam Mosseri post Patel quoted at length.

Mosseri's argument is that LLMs have made recommendation algorithms legible for the first time. Neural nets produce coordinates, not explanations. Now an LLM can look at those clusters and translate them into language — "here's why this video appeared" — which means you can talk back to the algorithm in plain English. That's the technical basis for Instagram's new explicit preference controls.

He then said something that I think most coverage understated: "A world where you are making personalized experiences on the fly is exciting in terms of agency but at the extreme starts to undermine shared experiences. If AI can generate entire apps and experiences that each of us wants, you and I might no longer share any sense of space."

The "filter bubble" conversation has been running since Eli Pariser named it in 2011. What got resolved in all that arguing: yes, personalization affects what you see; no, it doesn't fully collapse shared reality; the effects are real but unevenly distributed and context-dependent. What didn't get resolved: what happens when personalization stops being about content ranking and starts being about app architecture itself. Mosseri is describing the next version — not just different videos in your feed, but a structurally different app, with different affordances, generating different synthetic content, optimized for different goals.

The thing you can actually do right now: Instagram's new explicit preference controls are a real interface change, not just a marketing message. Use them deliberately, not reflexively. Tell the algorithm what you want more of, yes — but also notice what it had already decided you wanted before you told it anything. That gap is information about how the system had been modeling you, and it's worth knowing.

The more interesting question Mosseri doesn't answer: when bespoke app experiences are generated in real time, what does "you agreed to these terms of service" even mean? You agreed to an app that may no longer exist in the same form for anyone else.

The Vergecast hosts landed somewhere between fascination and unease, which is probably the right register. Patel noted that Mosseri, whatever you think of his decisions, "sees the thing" — he's not pretending the fragmentation risk doesn't exist. That self-awareness from a platform CEO is unusual enough to be worth crediting. It's also not the same as having a plan to address it.

---

Rachel "Rach" Kovacs is Buzzrag's cybersecurity and privacy correspondent.