Ripple effect: Google started turning Chrome, the world's most popular web browser, into an AI browser last year in response to threats from popular AI-native rivals such as OpenAI. Recent reports have uncovered that this transition includes silently installing a large cache of AI weights on an unknown but potentially significant number of devices.
Google Chrome users who have noticed unusual disk activity or unexplained drops in available storage should look for a folder called "OptGuideOnDeviceModel" inside their Chrome directory. It holds roughly 4GB of weights for Google's Gemini Nano LLM, downloaded by the browser without user consent.
Deleting the folder offers no lasting relief – Chrome will simply redownload it. On Windows 11, the folder resides at %LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel. It has also been confirmed on Apple Silicon and Ubuntu machines.
Uninstalling Chrome entirely is the most effective way to remove the weights. However, those who wish to continue using the browser might be able to disable the download by entering "chrome://flags" into the address bar, finding an item called "Enables optimization guide on device on Android," and selecting "Disabled" from the adjacent dropdown menu. This is also how users can determine whether their device is eligible for the feature.
According to Alexander Hanff, a computer scientist and lawyer who verified the behavior in the macOS kernel file system log, users have reported the automatic downloads for around a year. They might coincide with the introduction of AI features in Chrome last fall, such as writing assistance, AI summaries, and automatic browsing.
The downloads carry a notable irony: Chrome's most visible AI feature, the AI mode integrated into the address bar and Google Search, runs on Google's servers rather than the locally stored weights. The 4GB folder is only used for writing assistance and a handful of other features accessible several menus deep.
Given Chrome's billions of users, the total number of affected devices – and the bandwidth consumed – could be substantial. Hanff estimates that pushing 4GB to hundreds of millions or billions of devices would amount to several exabytes of data transferred, potentially generating between 6,000 and 60,000 metric tons of CO2.
However, determining the total number is difficult. This editor searched for the AI weights on two Windows 11 devices that are a few years old and found that both were ineligible, so Google might only be pushing the LLM to more modern machines.
Hanff has formally accused Google of violating European privacy regulations by compelling users to download a significant volume of data without their knowledge or consent.