Commotion about unsolicited AI file downloads in Chrome

News is currently circulating in the media that Google Chrome is simply wasting storage space by downloading a 4 GB AI model without any prompt. The supposedly uninhibitedly huge CO₂ footprint left by billions of affected browsers is also causing a stir.

The excitement apparently stems from a report by blogger “ThatPrivacyGuy”. He states that a 4 GB AI model named “weights.bin” is downloaded into the Google Chrome user profile directory without any prompt. The directory where it is located is reportedly called “OptGuideOnDeviceModel”. This is supposedly Google's Gemini Nano as an on-device LLM. ThatPrivacyGuy sees this as a violation of the GDPR, known internationally as GDPR.

Where is “weights.bin”?

Searching for the “weights.bin” file is unsuccessful on all accessible, tested systems with Google Chrome 148. A general, unprompted download cannot be reproduced so far. At this point, one might dismiss it as “much ado about nothing,” but it's not quite that simple.

Because the observation is not unfounded. Google actually documents this possibility for developers since the freshly released Chrome 148, which closes 127 security vulnerabilities. Under the heading “Prompt API,” Google explains that “Web developers get direct access to an AI language model provided by the browser on the device.” “The first implementation supports text, image, and audio inputs,” the developers explain, adding that it supports “various use cases, from generating image captions and performing visual searches to transcribing audio, classifying sound events, generating text based on specific instructions, and extracting information or insights from multimodal source material.” The Prompt API itself is documented on GitHub.

Browser-integrated, local AI functions

The Chrome browser receives integrated model management. Google writes that the integrated AI functions are based on Gemini Nano, which confirms ThatPrivacyGuy's assumption. Furthermore, Google's programmers document that the model is downloaded when needed. Last October, the developers thought it would be good to inform users about the time required for the download. The download is initiated by calling the AI API function create() if the model is downloadable: “The first download of the model is triggered by the first call to a *.create() function (e.g., Summarizer.create()) of any integrated AI API that depends on Gemini Nano.”

Google mentions another case where the local Gemini Nano model is downloaded: “Sometimes calling availability() can trigger the model download. This happens if the call is made shortly after starting a new user profile and fraud detection with Gemini Nano is active.” Those who enable the debugging pages intended for Chrome developers can gain insight into the current situation by calling “chrome://on-device-internals/”.

However, according to Google's documentation, settings for managing generative AI models should be found under “Settings” in the browser menu under “System.” According to a Google statement to Android Authority, the company has been rolling out the option to disable and remove the model via Chrome settings since February 2026 – however, it was not yet visible on the tested systems at the time of testing. This is likely because the feature is initially being rolled out in the US – here, the company has already been offering the Chrome browser with an integrated AI chatbot since last year.

The excitement about the CO₂ footprint from local AI files also seems to be exaggerated. Every single streamed Netflix movie generates this amount of data.

(dmk)