Google Chrome secretly installs a 4GB AI model on devices without consent

An investigation found that Chrome downloads the on-device model file weights.bin for Gemini Nano without user consent, and re-downloads it even after it's deleted.

The author claims to have confirmed the following flow by cross-referencing macOS's file event logs with Chrome's Local State, feature flags, and GoogleUpdater logs.

• On April 24, 2026, the OptGuideOnDeviceModel directory was created.
• On the same day, a background unpacker wrote weights.bin, manifest.json, verified_contents.json, and on_device_model_execution_config.pb.
• It was ultimately moved to OptGuideOnDeviceModel/2025.8.8.1141/weights.bin, with a total install time of 14 minutes 28 seconds.

Chrome's optimization_guide.on_device state retained component_version: "2025.8.8.1141" and hardware performance information, and OnDeviceModelBackgroundDownload was noted as enabled in ChromeFeatureState. This setting reportedly shows that the download can start before the user opts out via the settings screen.

The core issue is that Chrome's AI feature isn't a simple option, but rather a structure that automatically deploys a 4GB-class LLM model in the background. The post frames this as both a consent/transparency issue and a problem of large-scale disk/network resource consumption.