Apple's promise not to train AI on your data now has an exception

The change

Apple publishes a short page called Datasets used for Apple's generative AI systems and services, which describes the data it uses to train its AI models. Its section on personal information contained a clear promise: "Apple does not use our users' private personal data or user interactions when training our foundation models."

On September 9, 2026, Apple updated the page, and that sentence no longer ends there. It now reads: "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." Apart from the date, nothing else on the page changed. Venpo flagged the change on September 10, and the redline is on the public change page.

What changed

• The exception. The no-training promise now covers only people who have not chosen to help improve Apple's models.

• The same edit across Apple's AI pages. The same exception now appears in the Responsible AI principles on several of Apple's machine learning research pages, with a note: "Updated September 9, 2026, to reflect that people can now opt in to help improve Siri and Apple Intelligence."

• A new setting. Apple added Improve Siri & Apple Intelligence, which its support page says is "available for users of Siri AI." The page adds: "If you opt in to Improve Siri & Apple Intelligence, Apple can also use this data to train foundation models."

• One page still has the old sentence. At the time of publication, the Training Data section of Apple's research page on its third generation of foundation models still says "We do not use our users' private personal data or user interactions when training our foundation models," with no exception, while the Responsible AI section on the same page includes it.

What the new setting covers

Apple's Improve Siri & Apple Intelligence privacy page, dated September 14, spells out what opting in means:

• Apple "may store all of your interactions with Siri and Dictation," including the audio, transcripts of what you say and type, and the responses Siri gives.

• Conversations in the Siri app are included, and if you come back to a conversation, Apple may connect the random identifiers attached to it over time.

• If the ChatGPT extension is on, Apple may also store your requests and any conversation associated with them.

• This data may be used to train Apple's foundation models, which power Apple Intelligence and features in apps built on Apple's Foundation Models framework.

• It may be kept for up to two years, and the small subset that has been reviewed for longer. Graders may review samples, and only Apple employees review audio.

The older setting, Improve Siri & Dictation, remains for people who do not use Siri AI. Its privacy page describes using transcripts to fine-tune Siri, Search, Voice Control, Translate, and speech recognition models, and does not mention foundation models.

What stays the same

The new use of data depends on an opt-in. Apple asks during device setup, or the first time someone turns on or uses Siri, Dictation, or Translate, and the setting can be changed at any time under Settings, Privacy & Security, Analytics & Improvements.

Apple also keeps several protections in place. Opted-in data is linked to a random, rotating device identifier rather than to your Apple Account. Apple says it does not collect the photos or attachments you give Siri, the content of notifications Siri reads to you, or the information Siri searches on your device, and that it takes steps to remove phone numbers, email addresses, and content from messages and emails used in Siri's responses. Request history is not used to build a marketing profile and is never sold. The separate paragraph on aggregated Device Analytics data did not change.

Why this matters

An unconditional promise is easy to rely on. Before this update, anyone reviewing Apple's AI practices could point to one sentence and stop there. Now the answer depends on a setting that each person chooses, usually when they set up a device or first use Siri.

That means what you say or type to Siri, about your personal life or your work, can end up in Apple's training data because of one answer given in a setup screen. Apple's version is still an opt-in, which puts it ahead of services that turned training on by default, such as Lovable on its Free and Pro plans, or moved it behind an opt-out, as Replit did for Pro customers. It follows the same direction as Snyk adding an exception to its AI training promise: a flat no becomes a no with conditions. Knowing where your data can reach AI models is part of monitoring AI subprocessors.

Potential impact

For anyone who uses an iPhone or Mac, and for teams whose people do, the update raises four practical questions:

• Is Improve Siri & Apple Intelligence turned on for you, and do you remember choosing it?

• Do you ask Siri or the ChatGPT extension about things you would not want stored, such as health, money, or work details?

• If your team uses Apple devices for work, has anyone agreed whether this setting should be on?

• Who notices when a service you use every day adds an exception to a privacy promise?

One-sentence changes like this are what monitoring vendor terms and policies is for, because the rest of the page looks exactly the same.

How Venpo detected it

Venpo monitors Apple's legal pages, including its AI training data page, as part of continuous vendor risk monitoring. On September 10, the day after the update, it flagged the one sentence that changed and separated it from the new page date. The redline is on the Apple change page, and every monitored Apple document is listed on the Apple vendor profile. Our earlier case study covers another Apple change from the same week, when Apple gave itself the right to delete iCloud data over the storage limit.

Business outcome

Anyone tracking Apple learned the next day that its no-training promise had become conditional. That leaves time to check the setting on your own devices, and for teams to agree on a position before each person answers the prompt alone. The alternative is finding out later that the decision was already made in a setup screen.

Key takeaway

Apple did not drop its promise not to train AI on user data, it added eight words that make the promise depend on a setting each person chooses. When a vendor's clearest privacy sentence gains a condition, only automated monitoring of every wording change reliably catches it.