Google’s Android 15.2 focuses on privacy-preserving AI at the OS level

AI · 5 min read

Android 15.2 adds a system-level Model Store API that standardizes how apps download, validate, and sandbox on-device models. The store supports signed model bundles and a tiered permissions model that restricts sensitive sensors or data access during inference or training.

Complementing the Model Store is an Intent-aware Privacy Layer that manages AI feature permissions at a more granular level than existing runtime permissions. Users will be able to grant AI use for specific intents (e.g., summarization of owned documents) without opening blanket access to files or microphones.

Google also showed a demo of dynamic compute allocation where lower-precision servers handle heavy lifting and the device performs final personalization; this hybrid pattern intends to reduce network use while preserving a level of local control and explainability for results.