Meta logs employee mouse moves and clicks to train AI

Meta just found a new pool of AI training data, and it's coming directly from its own employees' fingers.

According to internal memos viewed by Reuters, Meta has begun installing software called the Model Capability Initiative (MCI) on computers of its U.S.-based staff. The tool tracks mouse movements, clicks, and keystrokes while workers use work apps and websites, and occasionally snaps screenshots of their displays.

Meta isn't spying to catch slackers. It wants to teach its AI agents how real people actually operate computers.

Turning daily work into agent training data

In a memo shared with the Meta SuperIntelligence Labs team, researchers explained that AI models still struggle with basic human computer interactions, such as picking items from dropdown menus or navigating complex keyboard shortcuts.

The fix is basically crowdsourcing inside the company: let thousands of staff go about their day, and let the software capture every click and scroll.

Meta CTO Andrew Bosworth outlined the broader vision under a newly rebranded program called the Agent Transformation Accelerator. Bosworth told staff the ultimate goal is a workplace where autonomous agents do the primary heavy lifting, while humans act as supervisors who direct and review. By watching where workers intervene, the models can figure out what went wrong and improve autonomously.

To push this through, Meta is restructuring internally. It wiped out several job classifications in favor of a universal "AI builder" title and created an Applied AI engineering team to funnel top developers into building agents that write and test software.

The office surveillance backlash

Meta spokesperson Andy Stone insisted the tracked data won't factor into performance reviews, adding that safeguards are in place for sensitive content. But Meta hasn't detailed what exactly gets filtered out.

Legal experts are already raising red flags. Yale law professor Ifeoma Ajunwa noted that while keystroke logging used to be reserved for misconduct investigations, Meta is now bringing gig-economy-style algorithmic surveillance to high-paid white-collar workers. In the U.S., federal law sets almost no limits on employee monitoring beyond basic notification.

Over in Europe, the tool faces a wall. Legal scholars note that electronic tracking of worker productivity is illegal in Italy, strictly curbed in Germany, and almost certainly conflicts with GDPR requirements.

It also lands at an awkward moment. Meta plans to cut roughly 10% of its global workforce starting May 20, following similar mass cuts across big tech. For employees, training an AI clone through their own daily muscle movements while layoff notices loom creates an inevitable morale crisis.

Why it matters

Building autonomous computer-use agents is the current holy grail in AI, but the bottleneck is authentic, fine-grained interaction data. Public web scrapers can't capture how an experienced engineer debugs a live system or how an analyst navigates clunky corporate software. By tapping its internal workforce, Meta gains a high-signal proprietary dataset that competitors simply don't have.

We think this move is technically clever but culturally reckless. Turning white-collar staff into passive data mules sets a dangerous precedent for corporate privacy. Even with promises that logs won't affect reviews, the psychological impact of knowing every mouse hesitation and backspace is recorded will inevitably chill how people work.

It also highlights how desperate frontier labs are for fresh tokens. First, companies began sifting through bankrupt startups' Slack archives and Jira boards; now Meta is harvesting the physical inputs of its own workforce. While it might give Meta a short-term boost in agent reliability, the looming GDPR showdown and the inevitable hit to employee trust make this look like a messy gamble.