Backflip: Apple now wants to train AI models with user data after all

Apple is abandoning its prior categorical promise not to use personal iPhone data to train AI: a privacy-policy change in the iOS 27 prerelease says Apple can store entire Siri and dictation interactions - including audio recordings, transcripts of typed and spoken requests, and Siri’s responses - to help develop Siri AI and in-house language models. The data will allegedly not be linked to Apple IDs, but some material may be reviewed by unspecified “review personnel” and audio will be listened to by Apple employees. Enrollment is handled via an opt-in prompt when activating Siri AI (with a subtle “Not now” option), and users can later disable it in Settings > Analysis & Improvements. The update was slated for public release on September 14, 2026.

The wording change also amends Apple’s “responsible AI” claims (removing an absolute ban on using private data unless users opt in) and undercuts assurances about Private Cloud Compute, which promised cloud-level performance with local-like privacy; iOS 27 also brings Google servers into that Private Cloud Compute mix. New Apple Watch features that provide 15-second conversation transcripts and a daily “Siri Recap” lean on Private Cloud Compute and thus on the updated data handling. The move revives concerns from a previous Siri privacy scandal in which contractor reviewers accessed intimate audio recordings.