Google says the on device AI model (Gemini Nano) inside Chrome powers real opt in features, but the install ran without consent and the Settings entry that named it was later edited.
Google did not ask before installing a 4-gigabyte artificial-intelligence model on Chrome users' computers. The on-device model, called Gemini Nano, lands in a folder named OptGuideOnDeviceModel and arrives without a system prompt, browser dialog, or notification when a computer meets Google's hardware thresholds. The complaint about that install path is now sharpened by a second decision: Google edited the Settings text that first named the on-device AI to users, narrowing the prior assurance that user data stayed on the device.
The install path is documented across the originating report, a Snopes fact check, and The Verge's coverage. Snopes independently verified the model file on three of six staff devices across macOS and Windows, and The Verge confirmed the same installer behavior on fresh installs.
Google announced the on-device Gemini Nano integration at its I/O 2024 developer keynote, pitched it as a way to run AI features locally without sending data to Google's servers, and continues to ship a Help Center page explaining how to manage it. The model file is called weights.bin; with supporting files, the on-device payload reaches around 4 GB on a fresh install and has reportedly accumulated past 12 GB on machines where older copies were not cleaned up after updates.
Android Authority and Ars Technica have both written that the model is not malware and is used to power opt-in browser features such as on-device translation, tab organization, and review summarization in developer tools. Google has publicly said so as well. The remaining question is consent, not capability.
That question opened wider when Swedish privacy researcher Alexander Hanff traced the install through macOS kernel log analysis. The install triggers on machines with a capable GPU tier, a multi-core CPU, at least 16 GB of RAM, and at least 22 GB of free storage. There is no opt-in dialog. Deleting weights.bin does not stop the next Chrome restart from re-downloading it; Chrome treats the deletion as a transient error and re-fetches the model.
The question got louder when Google edited the Settings text that first named the on-device AI to users. The That Privacy Guy blog and Decrypt both documented the language change. The prior wording included an assurance that on-device AI meant user data stayed on the device; the revised wording narrowed that claim. Whether the change reflects a product decision, a clarification, or both is something Google has not fully explained on the record.
On machines left running Chrome for a year, the on-device model has reportedly accumulated past 12 GB because older copies are not cleaned up after updates. For users on solid-state drives with constrained free space, that growth is not abstract. The hardware gate that triggered the install is also invisible: it does not surface in Settings, the flag menu, or any pre-install dialog, so a regular laptop owner learns the model is on disk only when they happen to look.
The unresolved piece is the one readers can act on. Chrome users on supported hardware can disable on-device AI features through the chrome://flags and chrome://settings paths Google documents, but disabling the features does not currently prevent the model from downloading. Independent researchers have asked Google to add a pre-install consent dialog, to clean up older model versions on update, and to restore explicit Settings-level language that names the model and explains the data flow. Google's response to those specific asks is the next data point worth watching.