Chrome Took 4GB and Didn't Ask. That's Not a Bug.

There's a file sitting on your Mac right now called weights.bin. If you run Chrome and your hardware clears a certain bar, it's already there. Four gigabytes. Google put it there. You weren't asked.

A writer covering the story for MacRumors flagged that Chrome has been downloading the file — which powers Google's on-device Gemini Nano model — without what they describe as explicit user consent. The model drives things like scam detection, autofill suggestions, and a writing assistant baked into the browser. The author credits a researcher named Alexander Hanff with noticing that Chrome installs it on any qualifying device, quietly, and without clear notice to the person whose drive just shrank.

The piece is useful. It tells you where the file lives and how to stop the bleeding. Read it if you need the fix.

But the fix isn't really the story.

What Google Stopped Pretending

The interesting thing isn't that this happened. It's that we're apparently still in the part of the cycle where we act surprised when it does.

For years, the industry framing around AI features has involved a lot of language about choice — opt-ins, toggles, settings menus, the careful choreography of consent. Google has been as loud as anyone about responsible deployment. And then Chrome reached into your storage and helped itself to four gigabytes of model weights because your machine met the minimum specs. No prompt. No explanation. Just a quiet transaction between Google's infrastructure team and your hard drive.

This is what it looks like when on-device AI stops being a feature and becomes an assumption. The browser didn't ask because, from Google's vantage point, there's nothing to ask. The model needs to live somewhere local to work locally. Your device qualifies. Therefore: your device hosts it. The logic is clean if you accept the premise — which is that your storage is now part of the cost of using the software, whether you've been told that or not.

The Infrastructure Admission

Here's what I keep turning over: on-device AI is genuinely different from cloud AI in one critical way that gets undersold. Cloud inference happens on someone else's servers. On-device inference happens on yours. That shift — which the industry loves to present as a privacy win, and sometimes it is — also means the compute cost and the storage cost migrate to you. Google gets to offer local processing as a feature while the overhead lands on your machine.

Four gigabytes is not catastrophic on a well-specced Mac. It's annoying on a machine running lean. It's a genuine problem on older hardware where that space was doing real work. And the size will grow. Models don't get smaller as capabilities expand.

What Hanff noticed, and what the MacRumors piece surfaces, is that the absence of disclosure isn't an oversight in the rollout — it's a signal about how Google has categorized this. Not a feature you adopt. Infrastructure you inherit.

The toggle exists. The article tells you how to find it. But the default was set without your input, and the default is always the confession.

Silent installs have a long history of being normalized right up until they aren't — and then someone writes the piece that tells you where the file lives.