Running AI inference on-device is often assumed to be privacy-preserving because it avoids sending data to remote servers. A new arXiv paper challenges that assumption for operating-system-integrated assistants, arguing that local computation does not, by itself, establish privacy.
The paper notes that an OS-level assistant can aggregate data across applications, retain derived state over time, and delegate actions to other components. Each of these behaviors can create privacy risks even when the underlying inference never leaves the device.
To address this, the authors present "feature-boundary auditing," an approach for examining where and how an assistant's features cross boundaries between applications, state, and actions. The paper is a replacement version (v2) of the arXiv preprint, though the abstract does not detail the auditing method's implementation.