Apple Reverses Course, Plans to Train AI with Siri User Data

Summary

Apple is changing a previous promise that personal user data and Siri interactions would never be used to train its foundation models. A privacy policy in the latest iOS 27 pre-release says Apple may store Siri and dictation interactions, including audio recordings, transcripts, typed and spoken requests, and Siri responses, to improve Siri AI and train its in-house language models. Apple says the data will not be linked to an Apple Account and that it will take steps to prevent personal data from being stored or used by Apple, but some material may be reviewed by unspecified personnel; the policy says only Apple employees will listen to audio. The new process is presented as an opt-in prompt shown when Siri AI is activated, although the dialog offers a visually inconspicuous “Not now” option and cannot be dismissed completely. Users can later disable the choice under Settings > Analysis & Improvements. Apple has also changed its responsible-AI wording by adding an exception for users who explicitly choose to help improve the models. The policy shift raises questions about Apple’s privacy claims for Private Cloud Compute, which supports Siri AI and other Apple Intelligence features, while Apple maintains that data is not stored long-term or shared with third parties. The change comes as new Apple Watch features are designed to listen continuously, provide a transcript of the previous 15 seconds, and generate a daily “Siri Recap” using Private Cloud Compute. Apple previously halted human review of Siri recordings after a privacy scandal involving contractors, later restoring it with opt-in consent. The iOS 27 update is scheduled for public release on Monday, September 14, according to the article.