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

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 obtained by Reuters and first reported by TechCrunch, Meta has begun deploying software on all US-based employees’ work computers under a program called the Model Capability Initiative, or MCI. The software records mouse movements, keystrokes, and periodic screenshots as employees go about their normal workday, across hundreds of applications and websites including Google Search, 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 been met with significant internal backlash, with employees flooding internal forums with questions about how to opt out — only to be told they cannot.

What the Model Capability Initiative Collects

The MCI software is installed on work computers used by Meta’s US-based workforce and activates during normal working hours. According to the internal memo, the program captures three primary categories of data:

Mouse movements and clicks — including which UI elements are targeted, how long users hover before clicking, and what sequences of actions are used to accomplish common tasks. This captures behavioral patterns that are difficult to infer from logs alone: how a human navigates a multi-step form, how they use keyboard shortcuts to cut time, which dropdown menus they access most frequently.

Keystrokes — the raw input that reveals not just what is typed but the cadence, error patterns, and correction behaviors that characterize human typing. This data is particularly valuable for training agents that need to interact with text fields, compose messages, or fill in forms across enterprise software.

Periodic screenshots — taken at intervals during the workday to provide ground-truth context for the other data streams. Screenshots allow the model to correlate keyboard and mouse activity with what was actually visible on screen at that moment, training the agent to understand the relationship between visual interface state and user action.

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 Search, LinkedIn, and Wikipedia are among the most prominent, but the list extends through the full range of productivity software, internal tools, and web services that employees use routinely.

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.

No Opt-Out: The Consent Controversy

The central controversy surrounding the MCI is not its technical logic but its consent architecture. Meta told employees that participation is mandatory and that there is no mechanism to opt out of the program. The company disclosed the program through an internal memo rather than through formal HR communications or policy amendments, and did not frame it as a change requiring employee agreement.

When the memo circulated, internal response was immediate and heated. Employees posted extensively in internal forums asking whether they could limit the data collection, what categories of activity would be recorded, how long the data would be retained, and what protections would prevent it from being used for purposes beyond the stated AI training objective. Multiple employees described feeling that a line had been crossed, with several noting that they routinely perform personal tasks — healthcare searches, financial queries, personal communications — on work computers, and that the boundary between professional and personal computer activity is not as clean in practice as Meta’s memo implied.

Meta’s official response drew a sharp distinction: the data collected through MCI is used “solely” for AI training and will not be used to evaluate employee performance. The company emphasized that participation in the program is a condition of using company-issued hardware, consistent with existing acceptable-use policies that grant employers broad rights to monitor work computer activity.

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.