Meta just told its US employees their keystrokes and mouse movements are now training data, and there's no opt-out.
The Summary
- Meta launched internal software that tracks employee keystrokes, mouse movements, and screen activity on work laptops to train AI models on how humans complete everyday computer tasks
- CTO Andrew Bosworth confirmed there is no option to opt out of the tracking on company-provided devices, sparking angry and shocked reactions from employees
- The program reveals a new training data frontier: companies turning their own workforce into unwitting data annotators to build AI agents that can replicate human computer use
The Signal
Meta's internal announcement, obtained by Business Insider, explains the company needs "real examples" of how people use computers to train AI agents. That means capturing keyboard shortcuts, dropdown menu selections, and the full spectrum of how humans navigate software interfaces. Fortune reported this represents a shift in how tech companies approach AI training data: mining how humans actually work, not just what they produce.
The employee response was swift and negative. The top-rated comment on Meta's internal workplace forum: "This makes me super uncomfortable. How do we opt out?" The angry-face emoji dominated reactions to the announcement.
"For agents to understand how people actually complete everyday tasks using computers, we need to train our models on real examples."
When Bosworth responded that "there is no option to opt out of this on your work provided laptop," employees hit back with crying, shocked, and angry emojis. The mandatory nature of the program cuts against years of tech industry rhetoric about user consent and data privacy. Meta employees, who build products that face constant scrutiny over data practices, now find themselves on the other side of that equation.
This isn't about productivity monitoring or catching slackers. Meta needs granular behavioral data to train computer-use agents. These AI models aim to automate tasks by mimicking human workflows:
- How someone navigates between applications
- Which keyboard shortcuts they use for efficiency
- How they interact with menus, forms, and interfaces
- The sequence of actions in completing common tasks
Meta says the software has privacy safeguards and limits tracking to work applications. But the backlash suggests employees see a line being crossed. When your employer captures every keystroke and cursor movement, the boundary between observation and surveillance blurs. The fact that it's for AI training, not performance reviews, doesn't seem to comfort anyone.
The Implication
This is the new data collection battleground for Web4. To build agents that truly automate knowledge work, companies need training data on the granular mechanics of how humans use computers. Your workforce is the most accessible source of that data. Expect more companies to follow Meta's lead, though probably with more polish on the internal comms.
If you work in tech, ask what data your employer is collecting now. If you're building AI agents, recognize that synthetic data and scraped content only get you so far. The real prize is behavioral data from actual workflows. Just know that harvesting it from your own people without genuine consent will cost you trust, even if it advances your models.