Meta Turns Employee Keystrokes Into AI Training Data

Summary: Reuters reported on April 21, 2026 that Meta was installing an in-house tool to capture employee mouse movements, clicks, keystrokes, and screen content as AI training data. The Model Capability Initiative covers United States employees and contingent workers on approved work applications, with no opt-out on work laptops, while Meta says data is used only for its stated purpose.

Reuters broke the story on April 21, 2026: Meta has been putting an in-house tool onto employee computers, one that captures mouse movements, click locations, and keystrokes, all of it destined to become training data for its AI models. Business Insider then got hold of the internal announcement and supplied the program’s name — the Model Capability Initiative (MCI) — which covers full-time employees and contingent workers in the United States.

This is more than a one-day story. Once operating a computer is itself a skill AI agents must acquire, the raw traces of how people interact with machines turn into a rare and coveted training input. Meta decided that the gathering would begin with its own workforce — and the internal Q&A stated outright that opting out is not on the table.

Inside the Model Capability Initiative

Here is what the tool logs:

Mouse movements, along with the location of every click
Keyboard input
On-screen content, used as context

The reach is limited. It runs only on work applications and URLs cleared in advance — Gmail, GChat, Metamate, and VSCode are the names the reporting cited — and it extends to computers alone, leaving phones out. A Meta spokesperson put it this way: “There are safeguards in place to protect sensitive content, and the data is not used for any other purpose.” One person familiar with the arrangement added that activity on Meta work devices was already under watch, and that employees were told as much when they were hired; read against that backdrop, MCI is a prolongation of standing policy more than a brand-new rule.

The Data Computer-Use Agents Cannot Do Without

The case Meta makes is technical. An agent meant to genuinely work a computer has to study authentic instances of people doing exactly that — expanding a dropdown menu, pressing a button, triggering a keyboard shortcut. In the spokesperson’s framing, models for everyday-task agents need “real examples of how people actually use them.” The announcement itself couples MCI with the debut of Muse Spark, the opening entry in a run of new models out of Meta Superintelligence Labs.

Stated plainly: text scraped from the web plus source code no longer cut it. An agent that has to drive an interface needs interface-interaction data, and that data is nearly absent from public corpora — the software everybody relies on daily is exactly the software whose interaction data is hardest to come by. Staff inside the company turn out to be the easiest source to tap.

Employee Pushback Meets a Blunt Answer

The dominant response in the thread was an angry-face emoji. The comment that collected the most likes said, “This makes me super uncomfortable. How do we opt out?” CTO Andrew Bosworth answered in that same thread: “There is no option to opt out of this on your work provided laptop.”

His reply is frank, and it marks the governance line with precision. The machines are company property, and consent to monitoring was signed at hiring — yet turning employees’ interaction behavior into model-training fuel assigns the pre-existing surveillance a new job. Whatever the legal footing, the unease is built into the arrangement: whatever the models call for is what gets gathered.

The Wider Pattern: Office Chatter Turned Training Data

Meta hardly hunts alone. As a Forbes report from April 16 described, the internal communications of startups that have folded or faded — Slack archives, Jira tickets — are being bought up and reprocessed into AI training data. TechCrunch ran the two stories side by side and kept a label worth repeating: a “troublesome privacy dimension” of the AI industry. With the public web growing thin, the industry’s attention has shifted to the digital exhaust of ordinary office days.

Takeaways for Enterprise AI Programs

Three takeaways. Data governance comes first: before a computer-use agent goes live, map where “interaction data” originates and what consent basis covers it — employees, customers, and contractors each fall under their own rules. Trust costs come second: Meta’s episode shows that a program can be legally airtight (device ownership, hiring-time disclosure) and still corrode trust inside the company the moment collection is compulsory — “no option to opt out” lands harder than any statute. Vendor due diligence comes third: when capability rides on acquiring interaction data, anyone sizing up agent products should add “where does the training data come from” to the checklist.

Computer-use agents sit near the top of 2026’s most hyped deployment bets. Through a single internal controversy, Meta has shown what they really cost: capability is paid for in interaction data — and interaction data comes from people.