Apple Turns the Old Siri Opt-In Into Model Training

iOS 27 now asks people who turn on Siri AI to share audio and text so Apple can train its foundation models. The prompt arrived with the public software on September 14, 2026, and it is optional.

The type of sample Apple already collected under Improve Siri and Dictation is largely the same. The new work for that sample is model training, which Apple’s own AI rules used to forbid in a sentence that ended with a period.

iOS 27 Asks Again, Even If You Already Said Yes

When Siri AI is turned on, the setup screen offers a choice labeled Improve Siri and Apple Intelligence. Apple’s on-screen copy asks for help improving foundation models and Siri by collecting text and audio from Siri, Dictation, and Translate.

THE SETUP CHOICES

• Share Audio and Text: Apple may store a sample of those interactions and use it to improve Siri and to train foundation models.
• Not Now: The assistant stays on, and the training permission stays off until you change it later.
• Older opt-in: People who had already agreed to Improve Siri and Dictation are asked again once Siri AI is enabled, so the extra training use is not inherited in silence.

The same release that rebuilt Siri as a chatbot also lengthened the consent screen that feeds the models. Not Now is the quiet exit, and it does not switch the new assistant off. You can reverse either choice later in Settings, under Privacy and Security, then Analytics and Improvements, using the Improve Siri and Apple Intelligence setting.

Apple Watch copies the iPhone or iPad that set it up. A Watch set up for a family member has the setting off by default, and it cannot be changed there. HomePod follows the iPhone or iPad used at setup. Apple TV matches that device unless you customize the box yourself.

The Sentence That Used to End With a Period

Apple’s public training-data notice, dated September 9, 2026, still says the company does not use private personal data or user interactions when it trains foundation models. The next clause is the change.

Apple does not use our users’ private personal data or user interactions when training our foundation models unless users explicitly choose to help improve them.

Apple, Datasets used for Apple’s generative AI systems, September 9, 2026

That exception now also sits in Apple’s responsible AI principles, which used to close the same idea without a loophole. The legal page still lists the rest of the diet: web pages crawled by Applebot, licensed and purchased sets, open-source material, user studies, and synthetic text and images. Apple says it has been collecting text for training since 2018 and images since 2020, and that the mix runs to trillions of data points.

The rewrite is the product of a rule Apple sold as a difference with chatbot makers that learn from live chats. Siri AI still runs on device and on Private Cloud Compute for ordinary answers. The new ask is for a stored sample, tied to a random device identifier, that can go into the next foundation models if you pick Share Audio and Text.

What Improve Siri and Apple Intelligence Collects

Apple’s support note is blunt about the bargain. Turn the setting on, and you agree to let Apple store and review a sample of audio recordings and transcripts. If you are on Siri AI, that sample can also train foundation models. If you are not, the older Improve Siri and Dictation label still covers Siri, Dictation, and Translate, without that extra training use.

WHERE THE DATA GOES

Related request data can include the category of the request, device specs, performance stats, and approximate location at the time. Apple says the bundle is tied to a random, device-generated identifier, not an Apple Account or email address, and that it is not used to build a marketing profile or sold. The company’s Siri privacy document still separates audio, which is not stored unless you opt in, from other Siri processing that happens to answer you.

Apple Already Ran This Drill in 2019

Human review of Siri audio is not a new idea at Apple. It is the scar the company has been managing since 2019, when graders working on Siri quality control were found to be listening to real requests, including accidental triggers.

THE OPT-IN APPLE ALREADY OWES

1. July 2019: Public reporting on Siri grading described contractors hearing private talk that users never meant to send.
2. August 28, 2019: Apple apologizes, saying it had not been fully living up to its own ideals, stops keeping Siri audio by default, and limits future audio listening to Apple employees for people who opt in.
3. January 8, 2025: Apple restates that it does not keep Siri audio unless a user opts in to help improve Siri, and that Siri data is not used for ads or sold.
4. September 9, 2026: The training-data notice adds that private user data and interactions stay out of foundation-model training unless users explicitly choose to help.
5. September 14, 2026: iOS 27 ships the Improve Siri and Apple Intelligence prompt, folding that exception into the old quality-control switch.

The 2019 fix was the opt-in. The 2026 move keeps the opt-in and assigns it a heavier job. Apple still says only employees hear audio, and that graders may see a subset of non-audio samples. The privacy fight from that year was about secret listening. This one is about whether “help improve Siri” now also means “help train the models that power Apple Intelligence.”

Reviewers Can Still Hear the Audio

Apple’s Improve Siri legal page, dated September 14, 2026, says Apple may retain this data for up to two years. A small subset of interactions that have been reviewed may be kept beyond two years for ongoing work on Siri, Dictation, Search, Translate, and related language features. The random identifier on that history rotates multiple times per hour.

If a request is selected for review, the question and Siri’s answer are graded for accuracy. Depending on the request, a reviewer may also see extra context that is directly relevant, such as contact names or apps on the device. Apple says only employees, under confidentiality rules, can access the audio. Graders can be used on non-audio samples.

That is a narrower room than the contractor era, and it is still a human ear on a voice clip. Device-side history of Siri and Dictation can be deleted in Settings, under Siri, then Siri and Dictation History. Turning Siri and Dictation off deletes the on-device transcripts used to personalize the assistant.

The Watch Recap Is a Separate Pipe

The new Apple Watch Series 12 and Ultra 4 add Audio Intelligence, and that is the feature people are mixing up with the iOS 27 training prompt. Apple’s own support page, published September 10, 2026, says Live Rewind and Siri Recap will be available in beta later this year, starting in English. They did not ship as finished tools on September 14.

WHAT THE WATCH ACTUALLY HOLDS

• Live Rewind: A double-press of the Digital Crown shows a text snippet of the last 15 seconds. The snippet clears about 30 seconds after the display dims unless you save it or tap Ask Siri. A chime, a full-screen animation, and a mic indicator fire even if the Watch is on silent.
• Siri Recap: Optional summaries of conversations, which auto-delete after 7 days if you do not save them. You can limit it by time or place, or toggle it from Control Center.
• Hardware cutout: Raw audio sits in a Secure Exclave on the S11 chip. Apple says the operating system, apps, the user, and Apple cannot read it, and that no recording file exists to hand over.
• Cloud step: Recap sends condensed text, designed to be less than half the original transcript, to Private Cloud Compute for a title and summary, then the PCC copy is discarded.

Recap is not offered to users under 13. On a Child Account, a parent or guardian has to turn Live Rewind and Siri Recap on for anyone under 18. Apple says the features do not label speakers, and that Recap tries to strip financial data, government IDs, and similar marks, with the usual warning that filters miss things.

Apple did not fold Watch conversation audio into the Improve Siri training description. The two systems can still look alike on a wrist that is listening for speech while the paired iPhone is asking for Siri clips. One is a later beta with a hardware vault and a seven-day note. The other is a stored sample that can last two years and, if you opt in, can train the models.

Training Stays Off Until You Share Audio and Text

Siri AI itself is a separate switch from the training ask. Apple Intelligence can be turned off under Apple Intelligence and Siri in Settings. The training sample is the Analytics and Improvements toggle, and it is the one that now names foundation models.

HOW TO CHANGE IT AFTER SETUP

• iPhone or iPad: Settings, Privacy and Security, Analytics and Improvements, then Improve Siri and Apple Intelligence or Improve Siri and Dictation.
• Mac: System Settings, Privacy and Security, Analytics and Improvements, same toggle.
• Apple Watch: Settings, General, Analytics and Improvements, then the matching switch, unless it is a family Watch that is locked off.
• HomePod: Home app, Accessory Settings, Analytics and Improvements, for the Home Organizer.

Apple’s software release notes for the September 14 platforms also warn that some server-side Apple Intelligence features have daily caps, and that extra access will be sold later. The training prompt is free labor on the other side of that ledger: a stored Siri sample if you agree, a paid ceiling on some cloud answers either way.

Siri AI is rolling out in beta on supported iPhones, iPads, Macs, Watches, and Vision Pro, and Apple says it will not be available in China while it works through rules there. On iPhone, iPad, and Watch, it is also held back in the EU at the start. The consent screen still appears where the assistant is actually on. The models get the clip only after Share Audio and Text, and they get nothing new from Recap until that Watch beta arrives and a user turns that on too.