Here's something you might not have expected.
Meta recently announced something, but without making a big fuss: they are going to start recording every keystroke and every mouse click from their employees — to train an AI agent that can automatically control computers.
Not external users, but employees. The people who work in Meta's offices every day.
What exactly are they collecting?
According to Meta, the data they plan to capture includes:
- Mouse movement trajectories
- Which buttons were pressed
- What was selected in dropdown menus
- Navigation paths across various applications
In plain terms: how you use your computer, they will record it all.
Meta's public explanation is that this data will be used to train an AI agent designed to help people complete everyday computer tasks.
The goal sounds reasonable enough — to build an AI that can actually use a computer, the best way is to watch how humans do it. But the issue is that they chose their own employees as the data source, rather than a specially designed data collection process.
Why this is more nuanced than it sounds
First, some industry context.
Over the past two years, AI companies have increasingly bet on computer control for the next generation of agents — not just chatbots, but systems that can open applications, fill out forms, click buttons, and get things done for you. This capability is called "computer use" or "computer-use agent."
The problem is that training such an agent requires vast amounts of real human-computer interaction data. Public datasets are of inconsistent quality, and collecting your own is expensive and slow.
Where is the most convenient data? Right there, inside the company, from employees.
This isn't just Meta's idea. Reuters reported that this is becoming an industry trend: Slack chat logs, Jira tickets, and employee operation logs are all being quietly considered by various companies for AI training.
Meta is just the first to say it out loud.
Is it reasonable?
Meta has given a standard answer: the collected data has protective measures and will not be used for other purposes.
But this answer sidesteps several real questions:
Which applications are being monitored? If you log into personal accounts or check personal emails on a company computer, will those actions be recorded?
Do employees have the right to refuse? Or is this already covered by the employment agreement, with no choice?
Can the data truly be isolated? It's easy to promise it won't be used for other purposes, but in the AI training pipeline, data flow is inherently difficult to track.
Meta's announcement did not answer any of these questions.
What's really worth watching
The fact that Meta wants to build a computer-use agent is not surprising in itself. OpenAI has Codex, Anthropic has Claude Computer Use, and Google is testing autonomous browser control. Everyone is racing on this track.
What's really worth watching is a bigger issue: the boundaries of AI training data are being redrawn.
Previously, the debate was about whether publicly available internet data could be used for training. Now people are asking whether employee behavioral data can be used. What will be next? User operation records within products? User conversation histories with AI?
These data points may already be authorized in terms of service agreements, but users — or employees — probably didn't realize they were signing up to be data labelers for AI when they agreed to the terms.
An AI that can replace humans in controlling computers needs human operation data to learn. That was originally a technical problem, but it is now becoming a more complex ethical one.
Meta's announcement is the beginning of this discussion, not the end.
Sources: CocoLoop, Meta will record employees's keystrokes and use it to train its AI models (TechCrunch)