Google Chrome has been quietly writing a 4GB artificial intelligence model file to users' hard drives without asking permission, a practice that became the subject of a formal European privacy complaint this week. The file, named weights.bin and stored in a folder called OptGuideOnDeviceModel, contains the parameters for Gemini Nano, Google's on-device language model that powers Chrome features such as text rewriting, scam warnings, page summarization, and tab organization.
The behavior was documented in detail by computer scientist and privacy lawyer Alexander Hanff, who reported that Chrome installs the model automatically whenever a device meets the hardware requirements, with no prompt, notification, or opt-in. If users delete the file to reclaim storage, Chrome re-downloads it on the next launch.
A formal ePrivacy complaint
Hanff has accused Google of violating the EU's ePrivacy Directive, which requires explicit consent before any software stores or accesses information on a subscriber's device. "This AI model is not an optional download," Hanff wrote, noting that Chrome installs it "without seeking explicit permission from the user." His complaint argues that the legal question hinges on the act of writing the file itself, not on whether Gemini Nano subsequently processes browsing data on-device or in the cloud.
Google's position is that Gemini Nano is designed to keep user data local. Because the model runs on the device's CPU or GPU, AI features can summarize pages or rewrite drafts without transmitting prompts to Google servers. The company began rolling out a settings toggle in February 2026 that lets users disable Chrome's built-in AI and remove the model, and once disabled the file is not re-downloaded.
How to stop the download
For users who want to prevent the install, the flag at chrome://flags labeled "Optimization Guide On Device Model" controls the behavior. Disabling it and deleting the OptGuideOnDeviceModel folder stops Chrome from re-fetching the weights. On Windows the folder lives under AppData/Local/Google/Chrome/User Data; macOS and Linux paths follow the equivalent profile structure.
A scaling problem, not just a privacy one
The story is also drawing attention from sustainability researchers. Chrome runs on more than three billion devices worldwide, and Hanff argues that downloading multi-gigabyte model files at that scale carries non-trivial bandwidth and storage costs that have never been disclosed to users or regulators. Even at 1% rollout, that translates to tens of petabytes of forced downloads.
Why this matters
The episode crystallizes a question every browser and operating system vendor will face as on-device AI becomes standard: does shipping a model count as shipping software, or is it a separate act that requires its own consent flow? European regulators have not yet ruled on Hanff's complaint, but the answer will shape how Apple, Microsoft, and Mozilla deploy their own local models in the coming year. For now, Chrome users who care about disk space or principle can disable the feature, but only if they know to look.
