Siri AI Is Rolling Out. Here's Where Your AI Requests Actually Go

On September 14, 2026, Apple began rolling out iOS 27 and a beta of Siri AI, a rebuilt assistant whose models Apple says were custom-built in collaboration with Google and its Gemini models and run both on the device and on Apple's Private Cloud Compute servers. For readers, that means the same assistant may handle a request on your phone, on a company's sealed server or through a third party such as ChatGPT, and where it goes decides what leaves your device, who keeps it and what you can switch off.

Key takeaways

Apple says many Apple Intelligence tasks run on device and complex ones go to Private Cloud Compute, where data isn't stored or made accessible to Apple. An iPhone report lists the requests sent there.
Pixel's Gemini Nano runs Call Notes and Scam Detection on the phone. Google's Private AI Compute, announced November 11, 2025, sends some tasks, such as Magic Cue on Pixel 10, to a sealed Google cloud. The Gemini app is a cloud service.
Samsung's Process data only on device switch blocks cloud processing for Galaxy AI, at the cost of features such as summaries and image editing.
On Copilot+ PCs, Recall and Click to Do run locally, but the Copilot app runs in the cloud. Since March 28, 2025, the Echo models that offered "Do Not Send Voice Recordings" no longer do; Alexa requests go to Amazon's cloud.
Local isn't automatically private, and "private cloud" rests on vendor commitments that outsiders can only partly check.

What it means for you

iPhone users (iPhone 15 Pro, iPhone 16 or later): Siri AI uses your phone and Apple's servers. It's a beta, rolling out in English first, and not initially available in the EU on iPhone, iPad or Apple Watch (EU Mac and Vision Pro users can use it). ChatGPT gets a request only if its extension is turned on.
Pixel and Samsung users: Pixel's on-device features are in the table. Samsung's switch covers Galaxy AI features; Google's Gemini app keeps chats under its own activity settings.
Windows users with a Copilot+ PC: Recall is off unless you opt in, and snapshots stay on the PC. The Copilot app sends what you type to Microsoft.
People handling health, legal or work information: cloud chat apps keep conversations for months, or until you delete them, by default, and some are read by human reviewers; see the retention table and follow your employer's rules for work data. This article explains vendor documentation and isn't legal or compliance advice.
People with older devices: without supported hardware, the on-device generative AI features described here aren't available, though Pixel 6 to 8 still run Scam Detection on the phone with older models. The cloud chat apps work on any device but send what you type to a server.

Where each assistant does the work

"On-device" means the AI model runs on your phone's or computer's own chip, so the data doesn't leave the device for that task. "Cloud" means it goes to a company's servers. A "private cloud" is one built so that, by the vendor's description, even the company can't see what it processes. The table reflects vendor documentation as of October 2026.

Assistant	On the device	Sealed company cloud	Ordinary cloud or third party	Main control
Apple Intelligence and Siri AI (iOS 27)	Many tasks, by Apple's description	Complex requests, on Private Cloud Compute	ChatGPT, only through its extension	Settings > Siri (ChatGPT extension); Apple Intelligence Report
Google Pixel	Gemini Nano: Call Notes, Scam Detection, Pixel Screenshots, Pixel Studio	Private AI Compute: Magic Cue on Pixel 10, Recorder summaries in more languages	Gemini app	Gemini Apps Activity (Keep Activity)
Samsung Galaxy AI	Most translations, Gallery edit suggestions	None described	Auto formatting, summarizing, auto corrections, image editing	Settings > Galaxy AI > Process data only on device
Windows on a Copilot+ PC	Recall; Click to Do analysis and text actions	None described	Copilot app; Click to Do "Ask Copilot" and web and visual search (Bing)	Settings > Privacy & security > Recall & snapshots, and Click to Do
Amazon Alexa and Alexa+	Wake-word detection	None described	Every request after the wake word	Alexa app: More > Alexa Privacy
ChatGPT, Claude, Gemini and Copilot apps	None for answering	None described	All chats, on the company's servers	Each app's data or privacy settings

Apple

Apple's September 14 announcement says the new Apple Foundation Models run on device and on Private Cloud Compute servers, and that Siri AI's Spotlight index and App Toolbox work entirely on device. Apple and Google's January 12, 2026 joint statement says Apple Intelligence will keep running on Apple devices and Private Cloud Compute.

The other platforms

Google's Pixel feature page says Gemini Nano on Pixel 9 and later lets Pixel Studio, Call Notes and Pixel Screenshots work without an internet or cellular connection. Its Private AI Compute announcement describes processing on Google's TPU chips in sealed "Titanium Intelligence Enclaves." Samsung's Galaxy AI support page says most features other than translation and Gallery edit suggestions need a network connection and use the cloud in some way.

Microsoft says Click to Do analyzes the screen locally and runs summarize and rewrite actions with Phi Silica, a small model on the PC's NPU (neural processing unit, a chip for AI work); those text actions need English, at least 10 selected words and a sign-in with a Microsoft or work account. Amazon's Alexa privacy FAQ says no audio goes to the cloud until the device detects the wake word or you press the Action button; after that, the recording is processed in Amazon's cloud. The last local option ended March 28, 2025, when, as TechCrunch reported, the Echo Dot (4th Gen), Echo Show 10 and Echo Show 15 dropped "Do Not Send Voice Recordings." Amazon said its generative AI features rely on its cloud.

Speed, capability and hardware

Vendors present the cloud as more capable: Apple says Private Cloud Compute uses larger, server-based models, and Google says some tasks exceed what on-device processing can do. Google says Pixel's Gemini Nano features, such as Call Notes, work without an internet or cellular connection. Battery claims come from vendors; Microsoft says Phi Silica uses very little power. Server-side Apple features carry daily usage limits.

Local AI needs specific hardware and space. Apple Intelligence requires an iPhone 15 Pro, iPhone 16 or later, iPhone Air, iPad mini (A17 Pro), MacBook Neo, or an M1 or later Mac or iPad. It takes up to 8 GB of storage, or up to 14 GB on models Apple lists, including the iPhone 17 Pro, iPhone 17 Pro Max and iPhone Air. Recall needs a Copilot+ PC with a 40 TOPS NPU (40 trillion operations per second), 16 GB of RAM, 8 logical processors and 256 GB of storage.

Illustrative example: on a 256 GB iPhone in the 14 GB tier, Apple Intelligence can take up to about 5.5% of advertised capacity (14 ÷ 256); on a 128 GB iPhone in the 8 GB tier, up to about 6.3% (8 ÷ 128). A Copilot+ PC with 60 GB free can turn Recall on (it needs 50 GB); snapshots pause below 25 GB, leaving about 35 GB of headroom (60 − 25). Windows also caps the space snapshots can use and deletes the oldest snapshots when the cap is reached.

How to tell which mode you are in

Try it offline. A feature that works in airplane mode runs on the device for that task; one that stops needs a server for at least part of it.
On iPhone, open Settings > Privacy & Security > Apple Intelligence Report, choose 15 minutes or 7 days, and export Apple_Intelligence_Report.json, a list of requests sent to Private Cloud Compute.
Watch for handoffs. Apple's ChatGPT guide says that with Siri AI nothing goes to ChatGPT until you ask for it. In Click to Do, "Search the web" and "Visual search with Bing" send your selection to Bing, and "Ask Copilot" hands it to Copilot.
Hit a limit. Apple's daily usage limits apply to features that rely on server-side models, so a limit message means a server was involved.
Flip the switch. On a Galaxy phone, the features that stop working under Process data only on device are the cloud ones.

Local isn't automatically private

On-device processing keeps the request off a server, but the results still live somewhere. Recall saves snapshots of your screen on the PC, including whatever other people send you. Microsoft's Recall privacy page says a sensitive-information filter, on by default, "helps" skip things like passwords and card numbers: a filter, not a guarantee.

Pixel's Call Notes stores call contents on the phone and doesn't share them with Google, so anyone who can get past your phone's lock screen could open them.

Results saved into synced apps follow that app's cloud rules. Under Apple's default standard data protection, iCloud Backup and Notes are encrypted, but Apple holds the keys. Turning on Advanced Data Protection makes them end-to-end encrypted, with keys only on your trusted devices.

What "private cloud" rests on

Apple says Private Cloud Compute uses data only to fulfill a request and doesn't retain it, and that your device checks a server cluster's identity and configuration (attestation) before sending anything. In October 2024 Apple released a security guide, a Virtual Research Environment that runs on a Mac with Apple silicon and 16 GB or more of memory, and source code for some components, with bounties of up to $1,000,000. It also publishes SOC 3 audit reports on the system that provisions those servers, the latest covering the period ending April 30, 2026. These let experts test Apple's claims, not inspect your own request.

Google says Private AI Compute data is "accessible only to you," not even to Google. An NCC Group assessment commissioned by Google spent 100 person-days on it between April and September 2025. It found a timing side channel in the relay that hides users' IP addresses, which could unmask users under particular conditions, with exploitability rated low. Its architecture review found the design stops a single insider from reaching user data, but noted that "it is logically unavoidable that Google ultimately has that power" as an organization. The phone apps that use the system were outside its scope. The ChatGPT handoff rests on a contract: Apple says that for users not signed in to ChatGPT, OpenAI must not store requests, unless the law requires it, or use them for training. Apple doesn't describe a way for users or researchers to check that, and signed-in users fall under their ChatGPT settings and OpenAI's own policies.

What the cloud apps keep

Once a request reaches an ordinary cloud service, its retention and training rules apply. As of October 2026:

Service	Default storage	Training and human review	How to limit it
Gemini apps	Auto-deleted after 18 months (3 or 36 months, or never, also offered)	With Keep Activity on, used to improve Google AI, including training; chats picked for human review are kept up to 3 years, disconnected from your account, even if you delete your activity	Keep Activity off or temporary chats: kept 72 hours and not used for training unless you send feedback, though still used for safety, including human review, per Google's Gemini Apps Privacy Hub
Microsoft Copilot	18 months	Used for training unless you opt out; not used for signed-out users, users under 18 or work accounts; some chats get human review, for example when a Code of Conduct violation is suspected, with no opt-out	Copilot privacy settings, per the Copilot privacy FAQ
Claude (Free, Pro, Max)	Kept in your history until you delete it; deleted chats leave back-end storage within 30 days	Training only if you allow it, then up to 5 years de-identified; flagged chats kept up to 2 years	Model improvement setting or Incognito chats, per Anthropic's retention page
ChatGPT	Saved in your chat history; turning off training doesn't delete or hide saved chats	Used for training unless "Improve the model for everyone" is off; the switch covers new conversations	Settings > Data controls, or Temporary Chat (not in history, not used for training, may be kept up to 30 days for safety), per OpenAI's data controls page (archived October 4, 2026)
Alexa	Recordings saved unless you choose 3 or 18 months or "Don't save"	Used to train Amazon's models; humans review "an extremely small sample"	Alexa Privacy; with "Don't save," transcripts still stay until the chat has been inactive 30 days

Frequently asked questions

Will Apple start charging for Siri AI?

Apple says server-based features, including Siri AI and Image Playground, have daily usage limits and that expanded access will be available for a fee in the future. It hasn't published a price or date.

Does the other person know when Call Notes is on?

Yes. Google says participants are notified when Call Notes is turned on during a call. The feature works on Pixel 9 and later, except A-series models, and in the US produces summaries in English only.

Does Pixel's Scam Detection send my calls to Google?

No, according to Google's Scam Detection help page. Processing happens on the phone, and no conversation audio or transcript is stored or sent to Google's servers. It's off by default and, in the US, works on Pixel 6 and later (with Gemini Nano from Pixel 9), with a SIM card from a supported country while you're in that country.

Can Microsoft or other users on my PC see Recall snapshots?

Microsoft says it can't access or view snapshots and that they aren't shared with other people signed in to Windows on the same device. Opening Recall or changing its settings requires Windows Hello, and you need face or fingerprint sign-in set up to use it at all. Recall also requires Device Encryption or BitLocker.

What do I lose with Alexa's "Don't save recordings"?

Amazon's page on deleting Alexa recordings says choosing "Don't save" deletes existing voice recordings and any Voice IDs on the account. Amazon's privacy FAQ says voice ID, which tells household members apart, then won't work. Requests are still sent to Amazon's cloud to be processed.

Settings to check this week

iPhone: in Settings > Siri, check whether the ChatGPT extension is on and signed in to an account, then set the Apple Intelligence Report to 7 days.
iPhone: consider Advanced Data Protection if you save AI summaries in Notes or use iCloud Backup.
Galaxy phone: try Settings > Galaxy AI > Process data only on device for a week.
Gemini app: review Keep Activity and the auto-delete period.
Copilot+ PC: in Settings > Privacy & security > Recall & snapshots, confirm whether snapshots are on and filter sensitive apps and sites; then check model training in Copilot's privacy settings.
Echo: in the Alexa app, open More > Alexa Privacy > Manage Your Alexa Data and set how long recordings are kept: 3 or 18 months, or "Don't save."
Chat apps: before pasting health or work records, turn off training or use a temporary or incognito chat.