Skip to content
FAQ

Meta Deploys Keystroke and Screen Tracking on Employee Computers to Train Agentic AI, Sparking Internal Backlash

Meta is rolling out tracking software called the 'Model Capability Initiative' on US-based employees' work computers, capturing mouse movements, keystrokes, and some screenshots across hundreds of apps and websites, including Google, LinkedIn, and Wikipedia. The company says the data is not used for any other purpose than training its AI models. Employees reportedly cannot opt out, and the rollout has drawn internal backlash over privacy.

5 min read

Meta is harvesting data from its own employees to train the next generation of its agentic AI models — and the workers being recorded had no say in the matter.

According to an internal memo reported first by Reuters, Meta is installing software on US-based employees’ work computers under a program called the Model Capability Initiative, or MCI. The software captures mouse movements, clicks, keystrokes, and some screenshots as employees do their daily work, across hundreds of applications and websites including Google, LinkedIn, and Wikipedia. The purpose, the company says, is to teach its AI agents how humans actually interact with computers so that those agents can eventually replicate and automate complex office tasks.

The rollout has drawn internal backlash, with employees calling it “dystopian” and many asking how to opt out, which is not possible.

What the Model Capability Initiative Collects

The MCI software is installed on work computers used by Meta’s US-based employees and runs on a designated list of work apps and websites. According to reports on the internal memo, the program captures mouse movements, clicks, keystrokes, and some screenshots.

Mouse movements and clicks, such as clicking buttons and navigating dropdown menus, are among the inputs Meta says its models need real examples of. The memo framed the effort as a way to improve models in areas where they struggle, such as navigating dropdown menus and using keyboard shortcuts.

Keystrokes are also captured, as part of the data fed into Meta’s AI training. This data is particularly valuable for training agents that need to interact with text fields, compose messages, or fill in forms across enterprise software.

Some screenshots are also taken. The memo said Meta needs to capture on-screen content as the context of what was being manipulated or interacted with.

The scope of tracked applications is broad. Meta has identified hundreds of applications and websites where it will capture this data. CNBC reported that Google, LinkedIn, and Wikipedia are among hundreds of websites and apps on the list, which also includes GitHub, Slack, Atlassian, Threads, and Manus and is still in flux.

The Technical Goal: Agentic AI That Mirrors Human Behavior

Meta’s stated objective is to build AI agents capable of performing complex, multi-step computer tasks on behalf of users — what the industry calls “agentic” AI or “computer use” AI. The challenge with such systems is not language understanding, which today’s large language models handle reasonably well, but procedural knowledge: the learned fluency with which humans navigate interfaces, manage workflows across multiple applications, and accomplish tasks through a non-linear series of mouse clicks, keyboard commands, and visual attention.

Current methods for training agentic AI systems typically rely on synthetic data — generated by having AI systems observe other AI systems, or by manually creating demonstrations — or on small-scale human annotation projects where contractors perform tasks in controlled environments. Both approaches have well-documented limitations. Synthetic data can encode systematic biases or fail to capture the full messy variety of real human computer use. Contractor annotation tends to be expensive, slow, and skewed toward simplified tasks that annotators can finish in well-defined sessions.

Meta’s Model Capability Initiative takes a different path: instrumenting its own knowledge worker population to generate massive quantities of authentic, high-fidelity behavioral data at essentially no marginal cost per data point once the infrastructure is in place. The employees going about their actual jobs represent exactly the kind of naturalistic, complex, long-horizon computer use that agentic AI systems need to learn — using the same tools, on the same corporate networks, for the same real-world purposes that enterprise AI agents will eventually need to replicate.

The central controversy surrounding the MCI is not its technical logic but its consent architecture. Employees asked how to opt out of the program, which is not possible. The program was disclosed in a memo to staff in a channel belonging to the Meta Superintelligence Labs team.

When the memo circulated, internal response was immediate and heated. Many employees responded with discomfort or with questions about how to opt out, and others raised concerns that the tool could expose sensitive data such as passwords and personal information. Multiple employees characterized the project as “dystopian” in internal messages, with concerns about exposure of personal information such as workers’ immigration status, health, or family members.

Meta’s response said safeguards protect sensitive content and the data is not used for any other purpose; a spokesperson also said it won’t be used to evaluate employee performance. Meta employees agree to the company monitoring their work devices when they are onboarded.

The Broader Implications

Meta is almost certainly not the only large technology company exploring employer-mediated human behavioral data as a training signal for agentic AI. The potential value of this data is enormous: a model trained on millions of hours of authentic knowledge-worker computer activity, across the full diversity of enterprise software, would have a significant empirical advantage over models trained on synthetic or annotated data.

But the MCI represents a significant escalation in what employees are implicitly signing up for when they accept company-issued hardware. Previous forms of workplace computer monitoring — logging of network traffic, recording of corporate communications, tracking of application usage time — were generally understood as security and productivity measures. A program that records every keystroke and takes periodic screenshots to train AI models that will eventually do the jobs being observed occupies different ethical territory, even if it is legally defensible under existing employment agreements.

The backlash inside Meta is an early signal of a tension that will intensify as more companies recognize the value of authentic behavioral data for agentic AI training. The question of what employees owe their employers in the form of behavioral data — and what disclosure, consent, and protections they are owed in return — has no settled legal or ethical answer. Meta, having moved first at scale, will likely be the company that forces that question into public view.

Whether regulators in the United States, Europe, or elsewhere will treat MCI as a new frontier of workplace data collection requiring new rules remains to be seen. What is clear is that the program marks a moment: the point at which AI training data hunger started reaching directly into the daily working lives of the people building these systems.

Correction (November 15, 2026): 18 sentences were corrected or removed to match the sources listed on this page.

Meta agentic AI employee monitoring AI training data privacy workplace AI
Share

Related Stories

JetBrains Survey: Over Half of All GitHub Code Is Now AI-Generated

A JetBrains survey of 10,000+ developers finds that 51% of code committed to GitHub is now AI-generated or substantially AI-assisted, with the AI coding tools market hitting $12.8B. GitHub Copilot retains the widest adoption but satisfaction growth has stalled, while Claude Code has surged to lead all tools on satisfaction metrics with 91% CSAT.

5 min read