Security & privacy

How to delete the 4gb file that chrome may have secretly installed on your device

At a glance:

  • Google Chrome has been quietly installing a 4GB AI model called Gemini Nano on some devices without user consent, as reported by computer scientist and lawyer Alexander Hanff.
  • The model performs on-device tasks like scam detection, text suggestions, and screenshot analysis, but only activates if hardware requirements are met.
  • Users can remove it by disabling "On-device AI" in Chrome settings or by uninstalling Chrome entirely; detailed steps for Mac and Windows are provided.

The silent install

Google has been quietly downloading a large AI model, Gemini Nano, onto devices running Chrome without asking or notifying users, according to Alexander Hanff, a Swedish computer scientist and lawyer known as That Privacy Guy. Hanff told CNET that the install happens without permission, and users won't know it's there unless they actively search for it. The file — roughly 4GB — resides in a folder called OptGuideOnDeviceModel within Chrome's user data directory, containing a file named weights.bin.

Gemini Nano is designed to run locally on smartphones and laptops rather than in the cloud, performing tasks such as detecting scam phone calls, helping write text messages, summarizing recordings, and analyzing Pixel phone screenshots. A Google spokesperson confirmed to CNET that the model will automatically uninstall if the device lacks sufficient resources like processing power, RAM, storage, or network bandwidth. The company also noted that in February 2025, it began rolling out a toggle in Chrome settings to let users disable and remove the model, after which it will no longer download or update.

How to check and remove Gemini Nano

If you want to find out whether the 4GB model is on your device, Hanff explained that Chrome does not surface it, so you must look for the folder manually. Here are the steps for both Mac and Windows.

On a Mac

  1. Open Finder by clicking the blue smiling face icon on the dock or pressing Command (⌘) + N on an empty desktop area.
  2. Click Go in the menu bar and hold the Option key so that Library appears in the dropdown menu. Click Library.
  3. Navigate to Application Support > Google > Chrome > Default.
  4. Check for a folder named OptGuideOnDeviceModel. If it exists and contains weights.bin, the model is installed.

To remove permanently on a Mac:

  • Open Chrome, click the three-dot menu in the top-right corner, then Settings > System and toggle off On-device AI.

On a Windows device

You can check using a Run Command or File Explorer.

Via Run Command:

  • Press the Windows key and R, paste %LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel and press Enter. Look for weights.bin.

Via File Explorer:

  • Navigate to C:\Users\[YourUsername]\AppData\Local\Google\Chrome\User Data\OptGuideOnDeviceModel and look for weights.bin.

To remove Gemini Nano on Windows:

  1. Open Chrome, go to Settings > System and toggle off On-device AI.
  2. While still in Chrome, type chrome://flags in the address bar and search for "optimization guide." Set "Enables optimization guide on device" to Disabled.
  3. Restart Chrome completely (use the menu to exit, not just close windows).
  4. Delete local files by navigating to \AppData\Local\Google\Chrome\User Data and deleting the OptGuideOnDeviceModel folder.

Why it matters

The silent installation raises serious privacy and legal concerns. Hanff suggested that the move might help Google cut costs by shifting AI inference off its own servers and onto users' hardware. "Running inference on users' own hardware allows them to push 'AI features' without the compute costs," Hanff told CNET.

But Hanff argued there could be regulatory consequences, particularly in Europe. He indicated that the install may breach the European Union's General Data Protection Regulation (GDPR) principles of lawfulness, fairness, and transparency. Additionally, given potential environmental impacts, Hanff said Google should have disclosed the installation under the Corporate Sustainability Reporting Directive (CSRD). "Google has given us every reason not to trust them with a history spanning two decades of global privacy violations at massive scale," Hanff told CNET. "So, I suspect they figured asking permission (what the law requires) would hinder their ability to push this model and, of course, whatever comes after it."

Editorial SiliconFeed is an automated feed: facts are checked against sources; copy is normalized and lightly edited for readers.

FAQ

What is Gemini Nano and why did Chrome install it without asking?
Gemini Nano is an on-device AI model that performs local tasks like scam detection, text assistance, and screenshot analysis. Google began secretly rolling it out in February 2025, installing the 4GB model on capable devices without explicit user consent, according to security researcher Alexander Hanff. A Google spokesperson told CNET that the model autoremoves if resources are low and that a toggle was added to Chrome settings in February to disable it.
How can I tell if Gemini Nano is installed on my Mac or Windows PC?
On a Mac, open Finder, go to Go > Library (hold Option key), then navigate to Application Support > Google > Chrome > Default and look for a folder named OptGuideOnDeviceModel containing weights.bin. On Windows, press Windows+R and paste %LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel to see if weights.bin exists. If either file is present, the model is installed.
What legal risks does Google face due to this silent install?
Alexander Hanff, a computer scientist and lawyer, argues the install may violate the EU's GDPR principles of lawfulness, fairness, and transparency because Google did not ask permission or notify users. He also pointed out that Google should have reported the environmental impact under the Corporate Sustainability Reporting Directive. Hanff cited Google's history of privacy violations and suggested the company pushed the model without consent to avoid hindering adoption.

More in the feed

Prepared by the editorial stack from public data and external sources.

Original article