Apple is reversing its public pledge never to train AI models on user data. Starting with iOS 27 (released September 14, 2026), the company will collect Siri audio, transcripts, and interaction logs to improve its language models and "Siri AI"—presented as opt-in during Siri activation, with a hard-to-dismiss dialog and a Settings toggle to disable later. The audio will be reviewed by Apple employees and contractors (unspecified "review personnel"). Critically, Apple removed the sentence "Your private personal data and interactions are never used to train our foundation models" from its responsible AI guidelines and added an exception: "unless you explicitly choose to help improve them." The timing is problematic: Apple's "Private Cloud Compute" infrastructure (positioned as privacy-preserving) now also routes data through Google servers, and new always-listening Apple Watches will generate constant conversation transcripts and summaries. This mirrors a 2019 scandal when contractors exposed they were hearing intimate details in accidental Siri recordings—Apple added opt-in then too. The source doesn't detail what "unspecified review personnel" means legally or how data retention actually works.