Google Chrome: a 4GB AI model installs secretly

⚡

Key Takeaways

1Google Chrome automatically downloaded a 4 GB AI model on certain devices without prior consent.

2The model, named Gemini Nano, performs tasks such as fraud detection and screenshot analysis.

3Users can uninstall Gemini Nano through Chrome settings or by removing the browser.

💡Why it matters — This practice raises questions about Google's compliance with GDPR in Europe, particularly regarding transparency and consent.

Google Chrome and the Automatic Download of Gemini Nano

Recent reports have highlighted that the Google Chrome browser has downloaded a 4 GB artificial intelligence model, known as Gemini Nano, without authorization on some users' devices. This model operates directly on devices such as smartphones and laptops, rather than in the cloud, raising concerns among privacy advocates about a potential violation of data protection laws.

An AI Model with Diverse Features

Gemini Nano is designed to perform various tasks, such as detecting fraudulent phone calls, assisting with text message composition, summarizing recordings, and analyzing screenshots from Pixel phones. Unlike the AI Mode feature in Chrome's address bar, which sends queries to Google Gemini servers, Gemini Nano operates locally on the device.

Installation and Removal of Gemini Nano

According to Alexander Hanff, a Swedish computer scientist and lawyer, Gemini Nano has been installed without authorization on certain Chrome browsers. The model only installs if the device meets specific hardware requirements. A Google spokesperson told CNET that Gemini Nano will automatically uninstall if the device lacks sufficient resources, such as processing power, RAM, storage space, or network bandwidth. For those wishing to remove it, they can uninstall Chrome or disable the "On-device AI" option in the browser settings.

In February, Google began rolling out the ability for users to easily disable and remove the model directly from Chrome's settings. However, users will not know when Gemini Nano was downloaded to their device, as Chrome does not highlight it and did not seek permission.

On a Mac

To check for the presence of Gemini Nano on a Mac, open Finder, go to the Library via the Go menu, and then navigate to Application Support > Google > Chrome > Default. If the OptGuideOnDeviceModel folder contains a file named weights.bin, the model is installed.

On a Windows Device

Windows users can check the installation by using the Run command with the path %LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel. The presence of the weights.bin file confirms the installation. Another method is to use File Explorer to navigate to C:\Users\[YourUsername]\AppData\Local\Google\Chrome\User Data\OptGuideOnDeviceModel and look for the weights.bin file. To get rid of the AI model on Windows, open Chrome, go to Settings > System, and disable On-device AI. While in Chrome, type chrome://flags in the address bar and search for Optimization Guide. Then, set Enables Optimization Guide on Device to Disabled. Finally, restart Chrome by completely closing it using the menu to quit, rather than just closing the windows. Delete local files by navigating to \AppData\Local\Google\Chrome\User Data and removing the OptGuideOnDeviceModel folder.

Potential Legal Consequences

Hanff expressed concerns about the legality of this practice in Europe, suggesting it could violate the EU General Data Protection Regulation (GDPR). He also mentioned that Google should have included this information in the Corporate Sustainability Reporting Directive, due to potential environmental impacts. Hanff stated that this initiative might aim to help Google reduce costs by shifting AI workloads from its own servers to the user's computer. "Running inference on users' hardware allows them to offer 'AI features' without the computational costs," he explained to CNET.

Hanff also emphasized a lack of trust in Google, citing a two-decade history of global privacy violations. "Google has given us every reason not to trust them," he said, suggesting that asking for permission, as required by law, could have hindered their ability to deploy this model and everything that follows.