Google $GOOGL Chrome has been downloading a 4GB AI model file onto user devices without consent, according to a report by Alexander Hanff, a Swedish computer scientist and lawyer who writes under the name "That Privacy Guy." Stored under the directory name OptGuideOnDeviceModel within Chrome's profile folders, the file — called weights.bin — holds the parameter data for Gemini Nano, Google's on-device large language model.
Chrome installs the file automatically on devices that meet certain hardware requirements, with no consent prompt and no checkbox in settings. Removing the file manually does not stick — Chrome treats the deletion as a temporary error and restores the full download at the next available opportunity. Hanff says that keeping the file gone requires either navigating to chrome://flags to shut off the relevant AI flag, applying enterprise policy controls that typical home users rarely have access to, or removing Chrome from the system altogether.
As a controlled test, Hanff set up a brand-new Chrome profile on a macOS machine and left it entirely untouched by any keyboard or mouse activity. Monitoring the macOS filesystem event log — which captures file activity independently of any application — he observed the model directory appear and the entire download finish in roughly 14 minutes, with a background tab sitting idle on a countdown timer throughout. Internal Chrome state files that Hanff examined indicated the browser had already assessed the machine's hardware and flagged it as a suitable candidate before any data transfer began.
In Hanff's view, the behavior runs afoul of Article 5(3) of the E.U. ePrivacy Directive, a provision that bars companies from placing data on a user's device unless that user has first given clear and informed agreement. He further argues the download conflicts with GDPR obligations around transparent and lawful data handling — and that because the browser operates normally without the model, Google cannot claim a necessity exemption to justify skipping consent.
In a statement to The Verge, a Google spokesperson said the model handles security tasks — including scam detection and developer API functions — entirely on the device, and that Chrome will remove it automatically should the hardware lack sufficient resources. "In February, we began rolling out the ability for users to easily turn off and remove the model directly in Chrome settings," the spokesperson said. "Once disabled, the model will no longer download or update."
Hanff also raises a separate concern about how the model is presented to users. Despite its prominent placement, the "AI Mode" pill in Chrome's address bar has no connection to Gemini Nano — anything typed there is sent off to Google's remote servers for processing. The local model is actually what drives background capabilities — things like assisted writing and tab organization — that the average Chrome user may never deliberately activate. Hanff argues this arrangement means users bear the storage and bandwidth cost of the local model while their most visible AI interactions still go to Google's cloud.
The environmental cost of the distribution is substantial, according to Hanff's calculations. Hanff's calculations put the energy cost of a single distribution round — covering 100 million to one billion devices — at 24 to 240 gigawatt-hours, with associated emissions of 6,000 to 60,000 tonnes of CO2 equivalent; neither figure accounts for future model refreshes or the repeat downloads that occur each time a user tries to remove the file. As Tom's Hardware notes, in many markets a 4GB transfer equals an entire monthly mobile data budget, meaning the uninvited download can translate directly into out-of-pocket costs for users on capped or metered connections.
Chrome holds more than 64% of the global browser market, with an estimated user base of between 3.45 billion and 3.83 billion people, according to Hanff's report.