Browsers quietly downloading AI models and how to stop it
A recent discovery involving Google Chrome has sparked debate: the browser can download a local AI model, roughly 4GB in size, without clearly notifying users. The file powers Gemini Nano, an on-device AI system designed to handle tasks like scam detection and developer features without sending data to the cloud.
Google insists this is actually a privacy win. By processing data locally, the browser avoids transmitting sensitive information online. But not everyone is convinced, especially when the download happens in the background and can reappear even after being deleted. The same behavior isn’t limited to Chrome. Microsoft Edge is also adopting similar on-device AI features, meaning more users could see unexpected storage and bandwidth usage.
To address this, Windows 11 now includes a new policy setting that gives users and administrators more control. Called GenAILocalFoundationalModelSettings, it allows you to block these AI models entirely or let them install as intended. For those who prefer control over convenience, setting the policy to “Disallowed” prevents the download and removes any existing model. It works for both Chrome and Edge, and can be configured through Group Policy or the Registry, even dynamically without restarting the browser.
To regain control, Windows 11 now supports a policy that lets you block these local AI models entirely.
Click on Run and type Regedit
Navigate to:
EDGE: HKEY_LOCAL_MACHINE\SOFTWARE\Policies\Microsoft\Edge
CHROME: HKEY_LOCAL_MACHINE\SOFTWARE\Policies\Google\Chrome
Right-click | New | DWORD (32-bit) Value
GenAILocalFoundationalModelSettings
Set the value to 1, which blocks AI model download, or 0 if you wish to allow automatic download.
Close the registry editor and restart the browser. Once set to 1, the browser won’t download the model and will remove it if it’s already there. The situation highlights a growing tension in modern software. Local AI can improve privacy and performance, but it also introduces new concerns from hidden downloads to resource usage and transparency.
