Apple rebuilt Siri on Google's Gemini, and a writer who quit it came back

Apple shipped two things at once this week that matter for how you design. A rebuilt Siri that people actually want to use, and its first foldable phone. Both give you fresh patterns to steal. Let me catch you up on what changed and what to bring to your next review.

The deep cut

• On-screen context is the new input. Siri now reads what a user is looking at, so Ivan Mehta asked about coffee bags and World Cup lineups.
• A trusted name beats a new logo. Apple rebuilt Siri on Google's Gemini and won back a writer who had given up on it.
• An always-on mic is a design decision, not a feature. Apple Watch now summarizes conversations, and The Verge's crew flagged the risk.

Siri stopped being a timer button

For years, Siri was for alarms and not much else. That changed. On iOS 27, Ivan Mehta says he is "actually using Siri again" for real, multi-step work: pull a delivery thread from Messages, draft a return note, save a Reddit roaster to a specific note. The engine under it is Google's Gemini, not the old ChatGPT patch job.

The lesson for your team is plain. Apple did not win trust back with a new logo or animation. It won by making the assistant reliable at chained requests that touch a user's own files, mail, and photos. If your product has an assistant that people ignore, the fix is not a fresh coat of paint. It is whether it can finish a real task.

The assistant that watches your screen

The bigger pattern shift is context. Siri now reads what is on screen and acts on it. Point the camera at a drink and ask what it is. Take a photo of a restaurant sign and ask for the best menu items. On the Mac, a keyboard shortcut lets you add the focused window to Siri's context or crosshair-select part of a page to ask about it.

This is the input model your users will start expecting. Not "type a prompt into a box," but "help me with the thing I am already looking at." Apple even wired Siri into the Camera app to split a bill or add a Wallet pass from what is in the frame. If your assistant makes people re-explain what is on their screen, you now look dated.

The blind spot Apple left open

Here is the catch worth flagging in a review. Ben Thompson argues Apple's biggest AI blind spot is its belief in the primacy of apps. Siri is smart, but it still hands off to individual apps rather than acting as one agent across everything. Mehta hit this himself: he struggled to get Siri to save notes to the right note until he "dumbed down" the note's name.

So the reference point is not perfect. When you copy the pattern, watch the handoff. A context-aware assistant that stumbles at the last step, saving to the wrong place, calling the wrong app, feels worse than no assistant at all. The magic dies at the seam.

The privacy line inside the wrist

Apple's coming Audio Intelligence puts an always-on microphone on the Apple Watch that can rewind the last 15 seconds of a conversation and recap what was said. The Verge crew broke down a little figuring out what it means that a watch now summarizes your talks. It is useful for someone deaf or hard of hearing. It is also a mic listening to people who never agreed to it.

If your roadmap includes ambient capture or summaries of real conversations, treat consent and controls as the core design work, not a settings afterthought. Apple put "user control, privacy, security" at the front of the pitch for a reason. The feature and the trust problem ship together.

Two screens, picked from Android

The iPhone Duo landed at $2,000 to $3,200, and Apple's move was to wait and copy the best parts of Android foldables rather than ship first. If you own multi-screen or resizable layouts, that patience is the point. Let others find the broken hinges and awkward states, then ship the version that feels considered. First is not the prize. Right is.

Three questions for your team

• Can our assistant finish a real multi-step task using a user's own data, or does it still just answer trivia? Test it on the messy stuff, not the demo.
• Where does our AI hand off to another app or screen, and does it break there? Map the seams the way Mehta hit the wrong-note problem.
• If we add any ambient capture, who is the person we are recording who never opted in, and what do we owe them?