Privacy researcher Alexander Hanff discovered THIS WEEK that Google Chrome silently installed a 4GB AI model on hundreds of millions of devices without user consent. The model, Gemini Nano, was deployed between April 20-29 through Chrome’s automatic update mechanism and cannot be permanently removed—manual deletion triggers automatic re-download. The discovery went viral on Hacker News on May 4-5 with over 1,100 points as developers expressed outrage over what appears to be a clear violation of EU privacy law.
The 4GB File You Didn’t Ask For
Chrome downloaded a file called weights.bin to your machine and stored it in the OptGuideOnDeviceModel directory inside your Chrome user profile. The file contains Gemini Nano, Google’s smallest on-device language model. Multiple users discovered the 4GB file when investigating unexpected disk space usage, only to find it automatically re-downloads after deletion.
The file paths are:
- Windows:
C:\Users\[USERNAME]\AppData\Local\Google\Chrome\User Data\OptGuideOnDeviceModel\weights.bin - macOS:
~/Library/Application Support/Google/Chrome/OptGuideOnDeviceModel/weights.bin - Linux:
~/.config/google-chrome/OptGuideOnDeviceModel/weights.bin
There’s no way to prevent the re-download through standard Chrome settings. Worse, Chrome doesn’t always clean up old model versions—some users on Hacker News reported 12GB or more accumulated storage from multiple versions.
This represents a fundamental shift from browsers as “user agents” (acting on your behalf) to “autonomous agents” (making decisions independently). You no longer have full control over what runs on your own device.
The EU Legal Violation
Chrome’s silent installation violates EU ePrivacy Directive Article 5(3), which explicitly requires “prior, freely-given, specific, informed, and unambiguous consent” before storing information on user devices. Chrome failed all four consent requirements: no prior consent, no option to decline, no specific disclosure, and no clear notification.
Article 5(3) is the legal basis behind cookie consent banners. The law states: “storing of information, or the gaining of access to information already stored, in the terminal equipment of a subscriber or user is only allowed on condition that the subscriber or user concerned has given his or her consent.”
Alexander Hanff documented using macOS filesystem logs that the installation occurred “even on Chrome profiles that have never received human input”—proving this was automatic, not user-initiated. This isn’t a gray area. It’s a clear-cut legal violation that could trigger regulatory enforcement from European data protection authorities.
The Hidden Climate Costs
At Chrome’s billion-device scale, the environmental impact of a single 4GB model push ranges between 6,000 and 60,000 tonnes of CO2-equivalent emissions. That’s roughly equivalent to driving 6,000-12,000 cars for a year.
The calculation is straightforward: if 500 million devices downloaded 4GB each, that’s 2 exabytes of data transferred globally. Using conservative estimates for network energy consumption (0.06 kWh per GB) and average grid carbon intensity (475g CO2/kWh), this equals approximately 28,000 tonnes of CO2.
Tech companies often position local AI processing as more environmentally friendly than cloud computing. This analysis reveals the hidden climate cost of deploying models at billion-device scale without consent. The irony is thick: Google violated environmental responsibility in the name of privacy-friendly features.
Why Google Did This (And Why Developers Are Angry)
Gemini Nano powers Chrome’s on-device AI features: “Help me write” text composition, scam detection for phishing attempts, tab group suggestions, page summaries, and smart paste. Google’s rationale is that local processing is more private than cloud alternatives since user data never leaves your device.
The model runs entirely on your machine through Google’s MediaPipe inference framework using TFLite format. When you use “Help me write,” no data is sent to Google servers—it’s processed locally.
Here’s the privacy paradox: Google violated privacy law to enable privacy features. The approach itself isn’t inherently wrong. Apple uses on-device AI models for similar features. But the deployment method—silent, unconsented, and permanent—erodes trust in Google’s privacy claims.
Developer reaction on Hacker News has been scathing. Top comments capture the sentiment: “At what point does Chrome stop being a user agent and start being Google’s agent on your machine?” and “The automatic re-download after deletion is the really egregious part. That’s not a bug—that’s designed behavior.”
This could be a turning point for Chrome’s dominance. Developer trust is critical for browser adoption, and Chrome’s “we know better than you” approach is reminiscent of Microsoft’s controversial Windows 10 update tactics.
Key Takeaways
- Check your Chrome directory for the OptGuideOnDeviceModel folder and 4GB weights.bin file.
- You can’t permanently remove it—Chrome automatically re-downloads the model after manual deletion.
- EU users should file complaints with national data protection authorities if this concerns you. This is a clear ePrivacy Directive violation.
- This sets precedent for browser autonomy versus user control. If Google faces no consequences, other browsers may follow.
- Consider alternatives if Google’s approach erodes your trust. Firefox, Brave, and other browsers haven’t deployed similar unconsented AI models.
At Chrome’s scale, consent matters. This isn’t about whether local AI is good or bad—it’s about who controls your device. The answer should be you, not your browser vendor.
