I read the headline about Meta paying to peek at how people use their latest AI model and sat with it for a minute before I figured out why it bothered me. Not the privacy angle, which everyone jumped on immediately. The economics. A company with more distribution than almost anyone on earth is buying permission to watch a few hundred sessions, because that’s the only way its assistant can find out what its users are actually doing. A browser agent gets that for free, and doesn’t have to send it anywhere.

Paid observation studies are old news in software. People volunteer, they get compensated, researchers watch and take notes. Nothing sinister.

But look at what’s being purchased. Context. Just context.

A telemetry program is a confession

A cloud assistant receives a text box and a submit event. That is the whole sensory apparatus. Everything else gets stripped off before the request leaves your laptop: the tab you had open, the thing you were reading four minutes ago, which of your six Jira tickets you’re actually stuck on, whether you’re staring at the admin view or the customer-facing one, the half-written reply sitting in a Gmail compose window that you abandoned because you couldn’t find the invoice number.

So the model answers a question that has been surgically separated from the situation that produced it. And then the vendor, noticing that its answers keep missing, goes and pays people to let it watch them work, so it can reconstruct from the outside the thing the product couldn’t observe from the inside.

That’s a hell of a workaround.

The thing is already on your screen

You’re logged in. The page is rendered. The data is right there in the DOM.

Nobody needs to buy it from you.

Why a browser agent skips the telemetry problem entirely

Dassi runs as a Chrome extension in the side panel, which means the context question has a boring answer: whatever tab you’re on. If you’re in Salesforce looking at an opportunity record, that’s the context. If you’re in Gmail with a thread open, that’s the context. There’s no connector to authorize, no OAuth scope to grant, no research program to enroll in, because the agent is already sitting inside the session you authenticated yourself, an hour ago, with your own password and your own 2FA prompt. This distinction matters more than it sounds like it should, and it’s the same reason cloud browser agents can’t see your tabs no matter how good the underlying model gets.

The page content goes to whichever model you’ve pointed it at, when you ask it to do something. Not continuously. Not to build a behavioral profile. Not to a research team.

I keep coming back to how weird the alternative is when you say it plainly. Meta has to run a program, recruit participants, handle consent, compensate people, and analyze recordings, all to approximate a signal that a side panel reads directly off the page. The extension isn’t smarter. It’s just standing in the right place. (You can grab it here if you want to see what I mean.)

Whose servers, though

This is where bring-your-own-key stops being a nerd preference and starts being the actual point.

If you run Dassi with your own Anthropic or OpenAI or Gemini key, the request goes from your browser to that provider. One hop. Dassi is not a middleman holding a copy, and there’s no vendor in the middle with a commercial reason to study your usage, because the vendor in the middle does not exist. I’ve written before about why AI companies don’t seem to trust each other with data, and the through-line is the same: every extra party in the routing path is a party with its own incentives about your sessions.

Meta’s program is consensual and compensated, which puts it well ahead of the industry norm of just harvesting quietly. Credit where it’s due. But the fact that context has become a thing worth paying cash for tells you exactly how starved cloud assistants are, and how much value is sitting unused in the browser window you already have open.

What I’d want to know

Whether the observed sessions actually helped. Whether watching someone use a chat box for an hour teaches you anything about the job they were trying to finish, or whether you just learn that they typed, waited, sighed, and rephrased.

I suspect it’s mostly the second one. You can’t recover the missing context by filming the person who’s missing it.

Either way, the next time a vendor offers you money to watch you work, you might mention that the tab was open the whole time.