Google Chrome silently installs a 4 GB AI model on your device without consent

· ai design · Source ↗

TLDR

  • Chrome downloads Gemini Nano weights (~4 GB, weights.bin in OptGuideOnDeviceModel) to user devices silently, re-downloading if deleted, with no opt-in UI.

Key Takeaways

  • The author verified the install on a fresh macOS audit profile that received zero human input; macOS .fseventsd kernel logs show the 4 GB drop happened in 14 minutes 28 seconds via Chrome’s own OnDeviceModelComponentInstaller, not GoogleUpdater.
  • Chrome’s OnDeviceModelBackgroundDownload feature flag and the settings UI revealing it are gated by the same rollout flag, meaning the download begins before the user has any UI to refuse it.
  • Chrome reads GPU and unified memory (vram_mb, performance_class) to qualify devices before any AI surface appears, stored in the profile’s Local State JSON.
  • At Chrome’s estimated 3.45-3.83 billion user base, the author calculates 6,000-60,000 tonnes of CO2-equivalent per model push and argues the behavior breaches ePrivacy Directive Article 5(3), GDPR Articles 5(1) and 25, and CSRD reporting thresholds.
  • The only durable removal paths are chrome://flags, enterprise policy, or uninstalling Chrome; manual deletion triggers re-download on the next eligible window.

Hacker News Comment Review

  • Core dispute: several commenters argued the model is just bundled software covered by Chrome’s update consent, analogous to a spellcheck dictionary; others countered that 4 GB delivered silently across trust boundaries, with no discoverable opt-out for home users, is categorically different.
  • Practical infrastructure pain is real: at least one commenter managing thousands of NFS-backed student home directories flagged unexpected 4 GB-per-user growth and repeated re-downloads on roaming profiles as an operational problem with no clean solution short of enterprise policy.
  • The author’s climate math drew skepticism; the energy-per-GB figure cited (Parssinen et al., 2018) is eight years old and likely overstates modern network intensity, weakening the environmental argument even if the consent argument stands.

Notable Comments

  • @davb: Flags concrete NFS and Windows lab-machine impact: thousands of students, AppData profiles that either accumulate 4 GB each or re-download on every clear cycle.
  • @jazzypants: Notes the same author previously labeled Claude Code “spyware” for writing Windows Registry keys, raising credibility questions about the framing.

Original | Discuss on HN