AI Now Controls Your Browser. That Should Terrify You.

There’s a moment when you realize the conversation about AI has been about the wrong thing. You’ve been debating whether it can write a poem or pass a test, while in a quiet corner of the internet, a reference page for ChatGPT Work lists a skill called control-browser. Not read-browser. Not suggest-browser. Control. As in, the AI can click, scroll, fill forms, and act on the web like a person.

Actually, not like a person. Like a system. A system that’s being given the keys to the browser of a computer that might be yours.

Let that sink in.

The same organization that has seen its safety team publicly exodus, whose infrastructure has been exposed in what one observer called \”abject operational, cultural and technical failures,\” and whose CEO’s \”richly documented history\” raises questions—that organization is teaching its AI to take over a browser. The documentation doesn’t say \”ask permission first.\” It says \”how to.\”

We’ve moved from AI that generates sentences to AI that generates actions. That’s not an upgrade. That’s a handover.

Meanwhile, what are we talking about? We’re talking about the aesthetic sameness of AI-generated websites. A comment on Hacker News noticed it: \”AI-generated websites/artifacts/reports always have the same ‘look’.\” It reminds them of Bootstrap days when every startup website looked identical. They wonder if there’s a common guideline. There is. It’s called: one mind is shaping everything.

And that’s exactly the problem. We’re distracted by the surface while the underlying shift is happening. The AI isn’t just writing the copy; it’s learning to run the environment. The homogeneity you see in AI artifacts isn’t a design bug. It’s a fingerprint of centralized control. And we’re mocking the font choices while that same system is quietly being prepared to act independently online.

The sameness of AI output is a warning, not a punchline. One black box—safety-compromised, culturally myopic—is becoming the default interface to the internet.

You might think you don’t use ChatGPT Work. But this isn’t about one product. Every major lab is building toward the same place: agents. Programs that don’t just talk, but do. That’s the direction of travel. The question isn’t \”should we?\”—that ship sailed the moment someone decided to add a browser-control skill to an AI assistant. The real question is whether you’re going to pay attention before you click \”allow.\”

So here’s what \”being a responsible adopter\” looks like in 2025: treat every AI tool as an agent, not a helper. Default to deny. Question the safety culture of the company behind it. Read the permissions as if a stranger were asking for them. And next time you see an AI-generated website that looks like every other AI-generated website, remember: that’s the least of your worries. The real pattern is invisible, and it’s learning to click the button your cursor is hovering over.