Meta sent an internal memo on Tuesday — June 2, 2026 — telling employees they can now pause the company’s workplace data collection for up to 30 minutes at a time. The tracking, part of a program called the Model Capability Initiative, logs keystrokes, mouse clicks, and screenshots of employee machines. The stated purpose: training Meta’s AI models.

The concession arrives after an April announcement of the program drew internal blowback. One employee told the BBC the prospect of having their own actions fed into the company’s models felt “very dystopian.” Fair enough.

But the 30-minute pause button — hailed in some corners as a privacy win — isn’t a concession. It’s an admission. If Meta’s leadership believed the data collection was purely about security and productivity, as the company has suggested, there would be no off switch at all. Nobody gets to pause a badge reader or a firewall. You can’t opt out of your building’s CCTV for half an hour because you’d like to have a private conversation in the hallway.

The fact that the pause exists tells you exactly what category this tracking belongs in. It’s not infrastructure. It’s extraction.

The Half-Hour Exception That Proves the Rule

There’s a concept in employment law called the “primary duty” test — it’s used to determine whether a worker is exempt from overtime. The idea is that you look at what someone actually does most of the time, not what the job description says.

Apply that logic here. If a company says it monitors computer activity for security, but then lets employees turn it off whenever they want — up to 30 minutes, enough for a coffee, a personal call, or a quick scroll through LinkedIn — then security isn’t the primary purpose. Security systems don’t have snooze buttons. Surveillance systems that want to look like something else do.

Meta employs roughly 70,000 people. Even assuming modest adoption of the pause feature, the company is implicitly conceding that thousands of hours of employee computer activity per week are not, in fact, security-critical. So what are they?

Training data, as the April memo made plain. Mouse trajectories. Typing rhythms. The granular, second-by-second movements of knowledge workers that look an awful lot like the movements an AI agent would need to mimic if it were going to replace one.

Whose Model Is Being Trained?

Let’s be precise about what’s happening here. Meta is building AI agents designed to perform digital tasks — navigating interfaces, filling forms, responding to messages. These agents need training data that shows how humans actually use computers. The cheapest, most legally uncomplicated source of that data is the workforce you already pay.

An engineer at a competing firm, sitting in a hotel bar at a conference last month, put it this way: “We looked at doing something similar. The legal team killed it. Not because it’s illegal — because the labor-law exposure depends entirely on whether you call it ‘productivity monitoring’ or ‘AI training data.’ Meta called it both, in writing, and now they’re scrambling.”

The memo matters. Written acknowledgments that employee monitoring feeds AI training are the kind of thing plaintiff’s attorneys highlight in future class actions. Tuesday’s 30-minute opt-out reads less like a benefit and more like a legal strategy — an attempt to retrofit consent onto a program that never asked for it in the first place.

The Real Privacy Fight Isn’t Employee vs. Employer

Much of the commentary around this story frames it as a labor issue: workers want privacy, employers want visibility, and the two are in tension. That’s true as far as it goes, but it misses the stranger dynamic underneath.

The data Meta collects from its own employees isn’t being used to manage them. It’s being used to build systems that manage everyone else’s employees — the AI agents Meta wants to license or sell. Meta’s workers are, in effect, an unpaid focus group for products that will surveil and automate workers at other companies.

That’s where the conventional take gets it wrong. The story isn’t “Big Tech monitors its own.” The story is “Big Tech is building the monitoring tools for every other employer, and the first test subjects are the people who happen to draw a paycheck from Menlo Park.”

A pause button — 30 minutes, no more — doesn’t change that. It just makes the whole arrangement a little easier to tolerate, one half-hour at a time.

Sources