On April 21, 2026, Reuters reported that Meta had begun deploying an internal tool on employees’ computers that records mouse movements, click locations, and keystrokes as training data for its AI models. Business Insider obtained the internal announcement and named the program: the Model Capability Initiative (MCI), covering US-based full-time employees and contingent workers.
This is not routine news-cycle material. When “using a computer” itself becomes a skill AI agents have to learn, raw traces of human-computer interaction turn into a scarce training resource. Meta chose to start collecting it from its own workforce — and made clear in the internal Q&A that there is no way out.
What the Model Capability Initiative Does
The MCI tool records:
- Mouse movements and click locations
- Keystrokes
- Screen content, for context
There are limits. The tool only runs on a pre-approved list of work applications and URLs — reporting names Gmail, GChat, Metamate, and VSCode — and it covers computers only, not phones. A Meta spokesperson said, “There are safeguards in place to protect sensitive content, and the data is not used for any other purpose.” A source familiar with the matter added that activity on Meta work devices was already monitored and disclosed at hiring, which makes MCI less a new policy than an extension of an existing one.
Why Computer-Use Agents Need This Data
Meta’s rationale is technical. To build agents that can actually operate a computer, models need real examples of human interaction — selecting from dropdowns, clicking buttons, using keyboard shortcuts. The spokesperson’s framing: models for everyday-task agents need “real examples of how people actually use them.” The internal announcement ties MCI to the launch of Muse Spark, the first in a series of new models from Meta Superintelligence Labs.
Put differently, web text and code are no longer enough. Agents that operate interfaces need interface-interaction data, and that data barely exists in public corpora — the most widely used software is precisely where data is hardest to get. Your own employees become the most convenient source.
Employee Backlash and the No-Opt-Out Answer
The most common reaction to the internal announcement was an angry-face emoji. The top-rated comment read, “This makes me super uncomfortable. How do we opt out?” CTO Andrew Bosworth replied in the thread: “There is no option to opt out of this on your work provided laptop.”
The reply is honest, and it draws the governance boundary with precision. The devices belong to the company, and monitoring consent was signed at hiring — but using employees’ interaction behavior to train models pushes existing surveillance into a new purpose. Whatever the legal footing, the discomfort is structural: whatever the models need is what gets collected.
The Bigger Trend: Work Chatter Becomes Training Data
Meta is not alone in hunting for fresh sources. A Forbes report from April 16 described how internal communications from defunct or aging startups — Slack archives, Jira tickets — are being bought up and converted into AI training data. TechCrunch paired the two stories under a label worth quoting: a “troublesome privacy dimension” of the AI industry. As public web data runs thin, the industry’s attention is turning to the digital exhaust of everyday office work.
What It Means for Enterprise AI Programs
Three implications. First, data governance: before deploying computer-use agents, companies should inventory where “interaction data” comes from and what consent basis covers it — employees, customers, and contractors each need their own rules. Second, trust costs: Meta’s case shows that even a legally defensible program (device ownership, hiring-time disclosure) erodes internal trust when collection is mandatory — “no option to opt out” is more persuasive than any statute. Third, vendor due diligence: when model capability depends on acquiring interaction data, buyers evaluating agent products should put “where does the training data come from” on the checklist.
Computer-use agents are one of the most hyped deployment directions of 2026. Meta has just demonstrated, through an internal controversy, what their cost structure looks like: the price of capability is interaction data — and the source of interaction data is people.
Sources
- Meta will record employees’ keystrokes and use it to train its AI models — TechCrunch
- Meta employees are up in arms over a mandatory program to train AI on their mouse movements and keystrokes — Business Insider
- Meta starts capturing employee mouse movements, keystrokes for AI training data — Reuters
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.