On 24 April 2026, Google Chrome reached into a researcher’s macOS profile, created a folder called OptGuideOnDeviceModel, and wrote a 4GB file called weights.bin to disk. The browser was idle at the time. Nobody had clicked on anything. Nobody had been asked. The whole thing took 14 minutes from start to finish.

That file is the weights (files required) for Gemini Nano, Google’s on-device AI & language model. The researcher who caught it was Alexander Hanff, a computer scientist and lawyer who happens to spend his time reading macOS kernel file system logs. If you delete the file, Chrome reinstalls it. If you never knew it was there, it sat there using up 4GB of your storage anyway.

This is happening inside the world’s most-used web browser, on machines belonging to ordinary people. Default behaviour. No beta flag, no developer mode. Across Chrome’s roughly 3.5 billion device install base, the math gets interesting fast. Every one of those devices that gets the push is another 4GB written to a disk that the user did not authorise.

I have been waiting for this story to land properly since it broke a couple of weeks ago, because it sits in exactly the same family as the iPhone age verification story I wrote about in March, and the broader pattern of “AI is being installed on your behalf, by people who think they know what is good for you” deserves a proper look.

What the model actually does

The model itself is genuinely useful in a few places. Gemini Nano powers Chrome’s “Help me write” feature, the on-device scam detection that flags suspicious pages, and the Summariser API that lets any website ask the browser to summarise content directly on the device.

What it does not power is the AI Mode that appears in the address bar and in Google Search. That all runs on Google’s servers. The 4GB sitting on your laptop is for the smaller, niche, in-browser features that almost nobody is using day to day.

That asymmetry matters. Most of the people who got the model would never knowingly use the features it powers. They opened Chrome, used it normally, and went to bed. Their hard drive was 4GB lighter when they woke up.

The privacy bit

Hanff, who has been involved in EU privacy law since the original ePrivacy Directive was being drafted, has formally accused Google of breaching it. The argument is straightforward. The ePrivacy Directive and GDPR both require user consent before storing data on user devices. Cookie banners exist because of this. The whole point of the regulation is that what gets written to your disk is your business, and platforms have to ask first.

A 4GB AI model is a long way past a cookie. It runs to roughly 750 photos worth of disk space, written silently while the browser is idle. If a website did this, it would be a regulatory event. Chrome doing it on Google’s behalf has been quietly defended as a “browser feature.”

It is a download. Calling it a browser feature does not change the legal question, which is whether the user agreed to the storage. Reasonable people will read the law differently, and the EU regulators will work out who is right in their own time. What is not arguable is that nobody asked.

The climate bit, which is bigger than people think

Google has not published a number for how many devices have received the model. Researchers writing about the rollout have estimated the climate cost at somewhere between 6,000 and 60,000 tonnes of CO2-equivalent, depending on the proportion of Chrome’s billion-plus active devices that get the push.

The lower end of that range is roughly the annual emissions of a small UK town. The upper end is closer to the annual emissions of 13,000 average UK households. For an unsolicited file download.

These are not numbers we usually attach to software updates because most software updates are kilobytes, not gigabytes. 4GB at this scale is the difference. It is the moment where “the install is free” stops being true for the planet, and starts being a meaningful externality. If the same 4GB had been pushed by an oil company, it would be on the front page of every paper in the country.

If you are running an AI-led business and you are starting to think about what your own AI hygiene looks like, the climate question is going to sit on the table sooner than you expect. Choose to push large models silently to a billion devices and you have made a climate choice. Whether you wanted to or not.

The pattern this story is part of

This is not the first time a platform has decided that quietly installing things on your device is acceptable. It is not even the first time this year. Apple’s iOS 26.4 update in March made every iPhone in the UK a child’s device by default until you proved otherwise, on the back of the Online Safety Act. Microsoft’s Copilot rollout has put AI features into Office without any meaningful opt-in step. The general direction of travel is that AI is becoming an infrastructure layer, and the people building the infrastructure increasingly behave as though consent is an obstacle, not a requirement.

There is a useful question to ask any time this happens. What is the smallest thing the platform could have done to make the rollout consensual? In Chrome’s case, a simple opt-in dialogue at the moment the user first uses one of the features Gemini Nano powers. “We need to download a 4GB AI model to enable this. OK or No thanks.” It would have been one extra click for a feature most people will never use. It was not asked. The model was simply written.

That is a deliberate design choice, not an oversight. Designs that ask for consent get lower opt-in rates than designs that do not. That is the whole reason the law exists. Convenience for the platform is not consent from the user.

What you can actually do

If you want the model gone, Google has confirmed you can disable and remove it. The path is Chrome → Settings → System → toggle “On-device AI” off. The model is not present on every device, so check first whether your machine has it. Open Finder, hit Cmd+Shift+G, and paste your Chrome profile path. If you see a folder called OptGuideOnDeviceModel, that is the one. The file inside it is weights.bin.

You can also keep an eye on what your browser is writing to disk if you are technically inclined. Tools like LuLu on macOS or osquery on either platform will show you outbound network requests and disk writes. None of this should be necessary in 2026, and that it is, for an AI model, is the bit that actually deserves attention.

Where this leaves us

I am not anti-AI. AI is one of the most useful things I have built my business around, and I think the next decade is going to be defined by businesses that learn to use it well.

What I am, and always have been, is in favour of the user knowing what is being done to their device. AI does not get a special pass on consent. It does not get a special pass on storage. It does not get a special pass on energy. The fact that it is the future does not make today’s behaviour acceptable.

If Chrome can quietly push 4GB to your disk for an AI feature you will probably never use, the next thing it pushes will be larger, and the one after that larger still. The line that gets drawn this year is the line for everything that comes after. That is the bit worth caring about, and the bit worth pushing back on, regardless of which side of the AI debate you are on.

The model can stay if Google asks. It can go if I say no. That is the principle. It really is that simple.

Scott Quilter | Co-Founder & Chief AI & Innovation Officer, Techosaurus LTD

Sources

• Google Chrome silently installs a 4GB AI model on your device without consent — Alexander Hanff, “That Privacy Guy!” (primary investigation)
• Google Chrome silently downloads 4GB Gemini Nano AI model — Quartz
• The climate costs are insane: why Chrome users are outraged over a forced 4GB Gemini AI update — TechRadar
• Google Chrome Is Silently Downloading a 4GB Gemini Nano AI Model to User Devices Without Consent — gHacks
• Guy finds Google Chrome is quietly installing a 4GB AI model on our devices — Cybernews
• Google Chrome Accused of Silently Installing 4GB AI Model on User Devices — HackRead
• Chrome Summariser API documentation — Google for Developers