Chrome silently downloads a 4 GB AI model without user consent

Chrome downloads a 4 GB local AI model onto every computer without asking. Google does not request permission and shows no notification. The browser simply presents it as a done deal.

What is happening in Chrome

Chrome automatically downloads the Gemini Nano model weights into the OptGuideOnDeviceModel folder. The file is called weights.bin and contains the weights of Google's local language model. It is 4 gigabytes in size. The browser downloads it with no consent dialog and no notification to the user.

If a user finds the file and deletes it manually, Chrome simply downloads it again on the next launch.

It is still unclear exactly how Google plans to use the model on the device. Possibly for analyzing text directly in the browser, without sending it to servers. But the fact that the file is downloaded without consent already violates users' expectations of what their browser may do.

History repeats itself

Two weeks ago, a similar scandal erupted around Claude Desktop from Anthropic. The application registered a Native Messaging module in seven Chromium-based browsers at once. That allowed Claude to write itself into the settings of other programs without the user's consent.

The scheme was simple: install Claude Desktop, and it automatically "re-registered" itself in Chrome, Edge, Brave and other browsers. There was no consent dialog and no visible way to opt out in Claude's own interface. If the user removed the entry manually, it reappeared on the next launch.

Now Google is using a similar approach, but a far more invasive one:

• The file is downloaded without asking the user
• There are no notifications in the browser interface
• There is no visible way to prevent it
• The file is restored automatically when deleted
• It affects billions of devices worldwide

The scale of the environmental harm

Here Google differs from Anthropic precisely in scale. Chrome is installed on roughly 3.5 billion devices worldwide. When a company decides to roll out 4 gigabytes of binary files to that entire audience, the environmental cost becomes genuinely material.

By estimates, a single wave of such downloads costs roughly 6-60 thousand tonnes of CO2 equivalent. Bandwidth, server electricity and energy on users' devices all add up.

For perspective: an average car produces roughly 4-5 tonnes of CO2 per year. So with this decision Google could generate as many emissions as 1.5 to 15 million cars driven for a full year.

The trust problem

Behind these two incidents lies a deeper problem: the boundaries of trust between vendors and users are growing ever blurrier. Companies claim the right to download and install things on other people's devices without explicit consent. They justify it with feature improvements, the need to run locally and better privacy. In reality, it leaves users without control over their own devices.

You cannot stop Chrome from downloading files, you cannot see what exactly is being downloaded, and you cannot turn it off in the interface. The file simply appears on your disk, taking up space and electricity.

What this means

The story shows an old problem in a new guise: large vendors claim the right to make decisions for their users. Privacy, environmental damage and control over one's own device all get pushed aside. When there are hundreds of such decisions and each affects billions of people, the problem takes on global scale.