Chrome's 4GB AI Model Is Installed On Millions Of PCs, But Is It Actually A Problem?

Google found itself at the center of public uproar when it was revealed that the company had been injecting a 4GB AI download onto the devices of Chrome users. This 4GB download mostly consists of the model weights (a weights.bin file) that power Gemini Nano, Google's on-device AI. Google did this without consent, a prompt, or even a notification –- users were rightfully upset. Google is increasingly becoming an AI-first company, and that core focus is shaping the default experience in its products, with Chrome at the top of the list.

This isn't exactly new; Google announced that Gemini Nano would be baked into the Chrome desktop client back at I/O 2024. Starting with Chrome v126, Google began rolling out on-device AI features using Gemini Nano. Google doubled down on this in 2025 with Chrome v140, expanding support for Gemini Nano to other devices to support more client-side AI features. The problem then is the same as it is now – Google was never clear about how this would look for the end user. Most of these announcements were geared towards developers, with little clarification on how and when these features would affect users. 

At this point, Chrome's 4GB AI model has been running on millions of PCs since 2024, in some capacity, probably unbeknownst to the majority of users. If you've used any AI features in Chrome over the last two years, there's a good chance these files were already on your device. Whether or not this is a problem depends on how you feel about privacy, user autonomy, and AI in general. I would argue that the AI model itself isn't actually a problem –- but the anti-privacy and lack of respect for users that Google is normalizing is.

Before anyone raises their pitchforks, let me be clear that I'm not here to defend Google. The sharp criticism Google has received over privacy, user choice, and its aggressive tracking of users is well known –- and in most cases, well earned. Google's "don't be evil" days ended a long time ago, and with it, its regard for asking users for their permission. If you're still using Google products, and most of us are, then you've likely come to terms with that. The case could be made that Google has been normalizing silent updates for years, delivering features to your browser whether you wanted them or not. Unlike Firefox, Google doesn't give you granular control over updates in Chrome (unless you're a system administrator), so we're all subject to Google's flexible definition of what constitutes an update.

A few years ago, the outrage over this would've been palpable. From a practical perspective in the age of the AI bubble, however, there isn't anything inherently evil about 4GB's worth of files for on-device AI if you find value in those features and can spare the storage. On-device AI that uses local processing is meant to be more private, but we also have no way of knowing if any metadata is being sent to Google's servers. Google also failed to clarify that its "AI Mode" in its search bar is not part of Gemini Nano, and relies on cloud computing. This is confusing at best, and deceptive at worst. If you use these AI features, you don't have to let perfect get in the way of good, but you should also expect to have a choice and a clear policy of how they work.

Alexander Hanff, also known as That Privacy Guy, is the one who broke the news that Google Chrome was installing a 4GB AI model on users' storage. Hanff is also the researcher who found similar behavior by Anthropic, in which it installed files in Chromium-based browsers without alerting users. In both cases, these types of files seem harder to remove than they are to install. With Gemini Nano, users were left guessing about how to remove it, and some reported that it would be reinstalled after it was removed. Google told CNET that the option to toggle the AI model on or off has existed since February, and after searching Google support docs, the company does indeed publish documentation for on-device AI models in Chrome, including how to turn AI off.

Google apparently determines what users get the download based on certain hardware requirements, but it doesn't seem to publish those requirements. I can verify that my desktop machine has the weights.bin file installed, while my MacBook Air does not, despite them running the same version of Chrome. Again, this is something Google should've been upfront about. Google also treats this as the new default in Chrome, without giving users the option to opt in, but instead forcing them to opt out by digging through Chrome settings or removing files from their computers. Regardless of whether you want the AI model or not, it's inexcusable.

According to Hanff, who is also a lawyer, Google's actions could face legal ramifications in Europe, where privacy laws are much more comprehensive and strict.

A number of companies have publicly embraced AI, even at the cost of thousands of layoffs, underscoring the harsh reality of a potential AI revolution. Consumers are less convinced, evidenced by the growing pushback against AI data centers in America. Among the environmental and economic concerns, there's also the question of ethics. Over the last several years, the collective of companies that make up Big Tech has repositioned itself to win the AI race, but how many of them can be trusted with the keys to AI? There's still significant mistrust of AI in general; a YouGov survey found that most Americans use AI in some way, yet they don't trust it.

A similar survey found that, among big tech firms, over 65 percent of Americans trust Google the most. So, for those that trust Google –- even if that trust is relative -– its aggressive push of AI is probably a non-issue. However, if you don't trust Google not to scrape data from your Gmail or invade your privacy, you certainly won't trust its behavior with AI. I find that Gemini Nano is less the problem and more the precedent Google has been establishing over the last several years. That it can leverage its dominant market share and captive users to deliver dubious features without warning or consent, treat them as opt-out by default, and then ask for forgiveness later – if at all.

From my vantage point, the backlash is less about a few gigabytes for an AI model, and more about Google making unilateral changes with potential privacy implications and shipping them to millions of users without notice. Zooming out even further, the bigger problem is the assumption that if you're using Chrome, Google has already decided it has your permission.