Chrome Installed 4GB of On-Device AI. What Google's Policies Authorize.

Google Chrome is downloading a 4GB AI model to your device without asking. The model is called Gemini Nano. It powers on-device features like text composition assistance and scam detection. If you delete it, Chrome downloads it again. There is no opt-out in Chrome's standard settings.

This story has generated significant attention and, predictably, a split between outrage and dismissal. Some coverage frames it as spyware. Others say it is nothing to worry about. Both miss the point.

We went to the policies. Here is what Google's own terms actually authorize, where the consent gap exists, and what this means for the 2 billion people who use Chrome.

What is Chrome downloading and why

Inside Chrome user profiles that meet hardware requirements, Chrome creates a directory called OptGuideOnDeviceModel containing a file called weights.bin. This file is approximately 4GB and contains the weights for Gemini Nano, Google's on-device language model.

Chrome uses Gemini Nano to power features including text composition assistance (marketed as "Help me write"), on-device scam detection in messages, and a Summarizer API that websites can call directly. These features are active by default in recent Chrome versions. The download happens silently when Chrome determines the device is eligible.

Reports of this behavior have appeared in community forums for over a year. The issue resurfaced recently when privacy researcher Alexander Hanff documented the full install chain using macOS filesystem logs, confirming that the download occurs even on Chrome profiles that have never received human input.

The part most coverage gets wrong

Gemini Nano is not malware. It is not spyware. It is not sending your data to Google.

This is an important distinction. When Chrome uses Gemini Nano to help you draft text or detect a scam, the processing happens entirely on your device. Your text never leaves your machine. Compared to cloud-based AI services where your prompts travel to remote servers, on-device inference is the more privacy-preserving approach.

If the only question were "is this model harmful to my privacy," the answer would be: no, it is actually better for your privacy than the alternative. The model processes your data locally instead of uploading it.

But that is not the only question.

The real issue is consent

The problem is not what Chrome installed. It is how.

No dialog box appeared asking whether you wanted a 4GB AI model on your machine. No notification told you it was downloading. No checkbox in Chrome Settings controls it. The download happens as part of Chrome's default behavior on eligible hardware.

More critically, if you find the file and delete it, Chrome downloads it again. Multiple independent reports confirm this re-download cycle across platforms. The browser re-downloads the file after manual deletion, effectively overriding the user's removal.

The only reliable ways to stop it are disabling AI features through chrome://flags, a developer interface most users do not know exists, or using Windows Registry edits and enterprise group policy tools designed for IT administrators. Regular users have no practical opt-out through Chrome's standard settings.

A 4GB download is not a routine browser update. Chrome updates are typically measured in megabytes. Four gigabytes is orders of magnitude larger. On metered connections, this consumes significant bandwidth without the user's knowledge. On devices with limited storage, it claims meaningful disk space.

Why this matters long-term

The Chrome Gemini Nano rollout reflects a broader shift in how software evolves after installation.

Modern platforms increasingly deploy new functionality silently through updates, introduce AI infrastructure incrementally, modify policies over time, and expand data processing capabilities after adoption.

For users, researchers, journalists, and regulators, the challenge is no longer simply reading terms once. The challenge is understanding what changed, when it changed, what new capabilities were introduced, and whether disclosures evolved alongside the technology itself.

This is why continuous policy and platform monitoring is becoming increasingly important.

What Google's terms actually say

ConductAtlas tracks Google's Terms of Service daily. The current terms grant Google broad authority to deliver automatic updates. We recently captured a notable change: Google's Terms of Service now state that services are provided with "reasonable skill and care," replacing the previous "as is" disclaimer. This shift implies a higher standard of conduct toward users, which may affect how the deployment is evaluated under applicable consumer protection frameworks.

Chrome's own Terms of Service, which we have now added to our daily monitoring, separately authorize automatic software updates to the browser. The legal question is whether a 4GB AI model download falls within what a user reasonably expects when they agree to keep their browser up to date.

A security patch is an automatic update. A version upgrade is an automatic update. Whether a 4GB AI model deployment falls within that same category is not explicitly addressed in the terms.

Where regulators may draw the line

Privacy researcher Alexander Hanff has argued that Chrome's behavior raises issues under multiple regulatory frameworks. The legal analysis centers on consent and proportionality, not on the model itself.

EU ePrivacy Directive, Article 5(3) requires prior informed consent before storing information on a user's device, with exceptions only for storage strictly necessary to provide a service the user explicitly requested. An AI model that powers default-on features the user never asked for is difficult to characterize as strictly necessary.

GDPR Article 5(1) requires processing to be lawful, fair, and transparent. Downloading software without notification, consent, or a practical opt-out raises questions under all three principles.

GDPR Article 25 requires data protection by design and by default. A system that installs capabilities automatically and re-installs them after user deletion raises questions under this requirement.

No regulatory authority has taken enforcement action on this specific behavior. But the pattern of silent installation without meaningful opt-out is the type of conduct that European data protection authorities have historically scrutinized.

The environmental dimension

At Chrome's scale of over 2 billion installations, even a partial rollout carries significant aggregate costs. Hanff estimated the carbon footprint of a single global model push at between 6,000 and 60,000 tonnes of CO2-equivalent emissions, depending on how many devices are reached. Each model weight refresh repeats the calculation.

For organizations subject to the EU Corporate Sustainability Reporting Directive, a supply chain dependency on Chrome now carries a potential Scope 3 environmental disclosure consideration that did not exist before this feature was pushed.

Timeline of events

The broader pattern ConductAtlas is tracking

This is not an isolated event. ConductAtlas has documented a recurring pattern across major platforms: capabilities expand under broad existing terms, often without prominent user-facing disclosure proportional to the change.

In the past month alone we have captured: OpenAI expanding data sharing to include marketing partners. GitHub explicitly permitting use of your code and outputs for AI training. Booking.com removing its "Do Not Sell" link from page footers. TikTok shifting its legal entity from a US joint venture to Singapore-based TikTok Pte. Ltd., reducing US-specific privacy protections.

Chrome's silent Gemini Nano installation fits this pattern. The terms authorize broad automatic updates. The implementation extends that authority to include capabilities the user may not have anticipated and cannot easily remove through standard settings. The gap between what the terms authorize and what the software implements raises questions about whether existing consent mechanisms are proportional to the scope of the deployment.

What You Can Do

Active monitoring

ConductAtlas continuously archives and tracks changes across platform policies, disclosures, and related legal documents.

Primary sources

• Google Chrome Terms of Service
• Chrome AI documentation
• Google Terms of Service
• Alexander Hanff's original research
• Technical removal guide (Pureinfotech)

View all Google policy documents tracked by ConductAtlas