Chrome's Hidden 4GB AI Model Sparks Consent Concerns

A ~4GB Model File Found in Chrome’s Folder

Reports have surfaced that a roughly 4GB AI model file was discovered inside Chrome’s user data folder.

The file is named weights.bin, and it holds the model weights for Gemini Nano, the on-device AI that runs locally on the machine. What made this a problem is that some users only found out about the file’s existence well after the fact. The controversy centers on whether users were adequately informed, at install time, about which feature uses this much space.

Privacy researcher Alexander Hanff said he examined the installation process on macOS, and similar cases have been reported on other operating systems. Whether the file gets installed depends on eligibility conditions and feature settings, so this doesn’t mean every Chrome user had the same file downloaded to their machine.

I personally find Chrome’s AI features convenient to use. Still, I think whether a feature is useful and whether users were sufficiently informed about it during installation are two separate questions.

Deleting a File Isn’t the Same as Turning Off a Feature

Let’s start with where the file lives and the conditions for reinstalling it.

The file in question sits under OptGuideOnDeviceModel inside Chrome’s user data folder. It’s not easy for an average user to look at that folder name and recognize it as space reserved for an AI model. I think this kind of information shouldn’t stay buried in an internal folder name — it should also be surfaced clearly in the settings screen.

weights.bin holds the model’s weights¹. Chrome checks conditions like supported operating system, storage space, and GPU or CPU availability before using the local model. The developer documentation states that the initial download requires at least 22GB of free space. This free-space threshold and the actual size of the model file itself are two different numbers. The model’s exact size can vary depending on updates.

Even if you delete the file yourself, it can be downloaded again later if the related AI feature is still turned on. That’s because deleting a stored file is not the same as turning off the feature that requires it. Users trying to reclaim disk space also need clear guidance on the setting that actually prevents reinstallation.

Google explained that it has shipped Gemini Nano in Chrome since 2024, and starting in February 2026 began rolling out a feature in settings that lets users turn off and remove the model. Google’s position is that turning off “on-device AI” in settings also stops the model from downloading and updating. At the time, some users reported that they couldn’t even see that option. Availability can differ depending on version, rollout status, and an organization’s management policies.

The benefits of local processing and installation consent are separate issues

Running a model on-device does have real advantages. It can process text without sending it to a server, and once downloaded, it can operate without a network connection.

Google explains that Chrome uses Gemini Nano for things like fraud detection and AI features for developers. Developers can use the built-in AI API to build features like text summarization. That said, not all of the AI built into Chrome runs on this model — different features use different processing methods.

I myself have found certain tasks easier thanks to AI help when using developer tools. But AI support in DevTools is a feature processed on Google’s servers. That’s exactly why my own convenient experience shouldn’t be described as evidence of Gemini Nano’s effect.

Even if local processing reduces data transmission, that doesn’t resolve the question of notice and choice when the file is installed.

Han-P argued that failing to obtain separate installation consent violates the EU’s ePrivacy Directive². Article 5(3) of the directive establishes the principle that consent is required when storing or accessing information on a device. However, there’s an exception for storage or access that is strictly necessary for a service explicitly requested by the user.

The mere fact that the file is larger than a cookie doesn’t determine whether the law was violated. What matters is which feature the user requested, whether installing the model was actually necessary for that feature, and what notice was given. The European Commission’s explanation also states that applying this exception requires judgment based on the specific facts of each case. We’re not at the stage where we can describe this installation as having been definitively ruled illegal or sanctioned.

Google explains that the local model is a product component for security features and developer APIs, and that it can be removed via settings. Setting aside the legal question, I think that if several gigabytes are being downloaded, users should be able to easily understand its purpose and the space it takes up.

AI Mode and Gemini Nano process data in different places

Even though they appear inside the same Chrome, different AI features can process data in different locations.

The AI Mode you reach from Google Search in the address bar is a server-side search feature. It’s not a function that answers using only your computer’s Gemini Nano. By contrast, features that call Gemini Nano through the built-in AI API run the model’s computation on-device.

Users should be able to see the following information clearly distinguished:

• Processing location: whether the feature computes on-device or sends requests to a server.
• Resource usage: the storage and network bandwidth needed to download the model, plus resource consumption during execution.
• Data handling: whether the feature or website collects and stores the inputs and outputs.

Only with this information can users weigh the benefits of local AI against the costs they personally bear.

It was also reported that Chrome 148’s on-device AI settings dropped the language stating that no data is sent to Google’s servers. Google explained that the way the model computes on-device hasn’t changed. The fact that the wording changed doesn’t, by itself, mean the model suddenly started running in the cloud.

According to Google’s explanation, a website using the model can see the inputs and outputs it handles, and that site’s own privacy policy applies. The model computing on-device and a website not sending data outward are two different guarantees. You also need to look at what data the webpage itself is processing.

Who Bears the Cost of Storage and Power

When evaluating on-device AI³, we need to look not just at features but at how it uses the device’s resources.

Downloading and running a model requires storage, network bandwidth, and power. To calculate the total cost, you’d need data on the actual number of devices with the feature installed, download counts, and usage time. You can’t simply take Chrome’s total user count and treat it as the number of installations.

What interests me is how that burden gets communicated to users, and whether they’re given a real choice.

In the cloud model, the provider’s servers do the computing; in the on-device model, the user’s device does. Each approach has trade-offs in cost, response speed, and data handling. A hybrid is also possible: tasks that a small model can handle run locally, while tasks needing a larger model get routed to a server.

The question is whether users actually received the information they needed to make that choice. It should be clear why a model is being installed, how large it is, which features need it, and how to turn it off if you don’t want it.

Installing a local model doesn’t suddenly turn your laptop into a server processing someone else’s AI workloads. What’s been confirmed here is the deployment of a local model for Chrome’s own features. The merits of the approach itself and the problems with how it was rolled out need to be discussed separately.

Users don’t need to know every model name or API structure. But they should be able to find out whether processing happens on-device or gets sent to a server, and how much storage it uses. To me, whether this explanation is adequately provided in the settings screen is central to trust.

Oswarld’s Lens

Honestly, I think on-device AI itself is a good direction. From my experience analyzing the architecture of various SaaS products while building GTM strategies, reducing dependence on the cloud is a choice worth considering in terms of both cost and data processing. That said, the actual benefit depends on device performance and the specific task.

What I found disappointing about this whole affair is that users only found out about the feature’s purpose and install size after the fact.

Something I’ve often noticed while doing data analysis is that even a feature built with good intentions can cost a company user trust if it’s applied in a way users didn’t expect. If Chrome had explained the model’s size and purpose before installing it, users could have judged for themselves whether they needed it.

I don’t know how many people would have agreed to the install, or what internal reasoning led Google to choose this approach. But making deployment easy and giving users a chance to choose shouldn’t be treated as mutually exclusive. Even if the goal is to distribute the model as widely as possible, that doesn’t justify skipping the explanation.

I actually use and enjoy Chrome’s AI features. I probably would have installed it even if asked first. That’s exactly why this bothers me. If a feature is genuinely useful, explaining its benefits and costs and letting users choose is the better way to earn their trust.

As other products start bundling AI models onto devices, similar problems are likely to crop up. Companies need to be much clearer about how they distinguish between features added via automatic updates and features users get to opt into separately.

Closing

What’s worth checking in Chrome right now is simple: which models are installed, which features are using them, and how to turn them off.

If the related AI feature is turned on, deleting the model file alone just means it gets downloaded again. To actually free up space, you need to check the feature settings too.

Local models can help reduce data transmission. But not all of Chrome’s AI features run locally, and running locally doesn’t mean a website’s own data processing is blocked either.

Beyond the question of legal obligation, what matters just as much is whether users could actually see the install size and purpose before choosing. I hope that disclosure becomes clearer inside the product itself.

If you use Chrome, type chrome://on-device-internals into your address bar and check right now which AI models are installed on your device.

If there’s an “On-device AI” entry under Settings > System, you can choose whether to use it. If your menu looks different, or if you’re on a browser managed by your company, check Google’s model management guide and your organization’s policy.

References & Further Reading

Primary sources

• Alexander Hanff, “Google Chrome Silently Installs a 4 GB AI Model on Your Device without Consent,” That Privacy Guy, May 4, 2026. : This is the original investigative report on the incident. It documents the forensic analysis process in detail.
• Jess Weatherbed, “Chrome’s AI Features May Be Hogging 4GB of Your Computer Storage,” The Verge, May 6, 2026. : A comprehensive report that includes Google’s official response.
• Snopes Fact Check, “Google Chrome May Have Silently Installed 4GB AI Model,” Snopes, May 8, 2026. : A fact-check article that verified the claim directly on Snopes staff devices.

Background

• “Chrome Deleted Its Own Privacy Promise for Sneaky On-Device AI,” Decrypt, May 7, 2026. : A follow-up report covering the changes to the settings text alongside Google’s explanation.
• “On-Device LLMs in 2026: What Changed, What Matters, What’s Next,” Edge AI and Vision Alliance, January 2026. : Useful for understanding the current technical state and trade-offs of on-device AI.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

Footnotes

1. Model weights: A file that stores, as numbers, everything an AI model learned during training. These are the values the model uses to process input. File size alone can’t tell you how good a model is. ↩

2. ePrivacy Directive: The EU’s electronic communications privacy regulation, which establishes the principle that storing or accessing information on a user’s device requires prior consent. There are exceptions to this consent requirement when storage or access is strictly necessary to provide a service the user has requested. The exact application depends on purpose and conditions of use. ↩

3. On-device AI: An approach where AI runs directly on a user’s device—smartphone, laptop, and so on—instead of on a cloud server. It can reduce the need to send data to a server for model computation. That said, whether an app or website is still collecting data is a separate question that needs its own verification, and this approach does consume the device’s storage and processing power. ↩