Google Chrome may be consuming extra storage on your machine. Reports indicate that the browser has been silently downloading a 4 GB AI model to the hard drives of some users without their consent. This isn’t the first instance of Google subtly modifying users’ devices, and privacy experts warn that the behavior could breach data‑protection regulations.
The file in question is Gemini Nano, an on‑device AI model designed to run locally on smartphones and laptops rather than in the cloud. Alexander Hanff, a Swedish computer scientist and attorney who goes by “That Privacy Guy,” says the model has been installed on certain Chrome browsers without permission, and users receive no notification when the download occurs.
According to Hanff, Gemini Nano will only be installed on devices that satisfy its hardware requirements, though the exact number of affected users remains unknown.
Gemini Nano can perform functions such as identifying scam calls, drafting text messages, summarising recordings, and analysing Pixel phone screenshots. It should not be confused with the AI Mode button in the address bar; when AI Mode is used, queries are sent to Google Gemini servers, not to the on‑device model.
A Google spokesperson told Gfaloe that Gemini Nano will automatically remove itself if the device lacks sufficient resources—CPU, memory, storage, or network bandwidth.
“In February we started rolling out an option for users to easily disable and delete the model directly from Chrome settings,” the spokesperson added. “Once turned off, the model will no longer download or receive updates.”
More details about on‑device generative AI models in Chrome are available on Google’s support page.
How to remove the AI model
If you suspect the 4 GB model is on your device, first verify its presence.
Hanff notes that Chrome users won’t see Gemini Nano unless they specifically search for it, because “Chrome didn’t ask” and “Chrome doesn’t surface it.” The simplest way to eliminate Gemini Nano is to uninstall Chrome.
On a Mac
- Open Finder by clicking the blue smiling‑face icon on the far left of the dock.
- Choose Go from the top menu, hold the Option key, and select Library when it appears.
- Navigate to Application Support > Google > Chrome > Default and look for a folder named OptGuideOnDeviceModel. If it exists and contains a file called weights.bin, the model is installed.
- To delete it, open Chrome, click the three‑dot menu → Settings → System, and switch off On‑device AI.
On Windows
- Press Win + R, paste %LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel, and hit Enter. If the folder opens, check for weights.bin.
- Alternatively, open File Explorer and go to C:\Users\[YourUsername]\AppData\Local\Google\Chrome\User Data\OptGuideOnDeviceModel to look for the same file.
- To remove the model, open Chrome, go to Settings → System, and toggle off On‑device AI. Then type chrome://flags in the address bar, search for “Optimization Guide,” and set Enable Optimization Guide on Device to Disabled.
- Restart Chrome by fully exiting the application (use the menu to quit, not just closing the window).
- Finally, delete the local files by navigating to \AppData\Local\Google\Chrome\User Data and removing the OptGuideOnDeviceModel folder.
Why it matters
Hanff believes Google’s push may be aimed at cutting costs by shifting AI processing from its own servers to users’ hardware.
“Running inference on users’ own devices lets Google add ‘AI features’ without bearing the compute expenses,” he told Gfaloe.
Inference—the stage where the model actually generates responses—occurs locally, which can affect device performance, battery life, and storage usage.
Hanff also warns of potential legal issues, especially in Europe. He suggests that silently installing Gemini Nano could breach the EU’s GDPR principles of lawfulness, fairness and transparency, and that, given possible environmental impacts, Google should have disclosed the rollout under the Corporate Sustainability Reporting Directive.
“Google has given us every reason not to trust a company with a two‑decade history of large‑scale privacy violations,” Hanff said to Gfaloe. “I suspect they avoided seeking permission because it would have hampered their ability to push this model and whatever follows.”