An internal memo that lit the fuse. On April 21, Meta announced to its U.S. employees the rollout of a new tool named Model Capability Initiative (MCI), capable of recording their mouse movements, clicks and keystrokes, as well as taking occasional screenshots of their screens.
The stated objective: to feed the company’s AI models with real-world behavioral data.
According to internal communications obtained by Reuters and Business Insider, the program is limited to a predefined list of professional applications — Gmail, GChat, Metamate (Meta’s internal AI assistant) or VSCode — and only applies to computers provided by the company, not phones. It covers full-time staff as well as contractors based in the United States.
Model Capability Initiative (MCI) limited to the USA
The internal reception was, to say the least, cool. According to Business Insider, which reviewed the threads on Meta’s internal messaging platform, the most-liked comment in response to the official announcement read: “This makes me really uncomfortable. How do I opt out?” The angry-face emoji was the most common reaction to the original post.
The response from chief technology officer Andrew Bosworth dashed hopes: “There is no opt-out option on your work computer,” he wrote in the thread, triggering a flood of shocked and tearful emojis in return.
A Meta employee, who spoke to the BBC on condition of anonymity, summed up the mood: “This company has become obsessed with AI,” adding that seeing every action on a computer used to train AI models, in a context of anticipated layoffs, felt “very dystopian.” A former employee, also cited by the BBC, saw it as “the latest way they’re forcing AI on everyone.”
On the management side, the stance is pragmatic. “If we build agents to help people perform everyday computing tasks, our models need real examples of how people actually use them,” said Andy Stone, the group’s spokesperson, cited by Reuters. The data collected would be used “for no other purpose,” he added, noting that there are “guardrails” in place to protect sensitive content, without providing further details.
A radical transformation of work
This rollout sits within a broader strategy. Rebranded internally as Agent Transformation Accelerator (ATA), the initiative, according to a memo from Andrew Bosworth consulted by Reuters, aims to “build a vision in which our agents do most of the work and our role is to lead, review, and help them improve.”
Meta has also announced the creation of a new engineering team, Applied AI (AAI), charged with sharpening the coding capabilities of its models and designing autonomous agents capable of building, testing, and delivering future products. The group’s top-performing software engineers have been transferred there.
The company, which has already undertaken job cuts this year, plans to lay off 10% of its global workforce starting on May 20, according to Reuters, with potential further reductions later in the year. A sign of the contraction, Meta’s job site listed around 800 roles in March; it now lists only seven today, according to the BBC.
Financially, the ambition matches the upheavals announced. Mark Zuckerberg plans to invest around $140 billion in AI in 2026, nearly double the previous year. In 2025, Meta had already taken a majority stake in Scale AI for $14 billion. In January, the CEO predicted that 2026 would be “the year AI would radically change the way we work.”
A legal vacuum in the United States, a wall in Europe
Legally, the situation differs markedly across the Atlantic. “On the American side, at the federal level, there is no limit to worker surveillance,” notes Ifeoma Ajunwa, a law professor at Yale University, interviewed by Reuters. State laws at best require prior notice to employees.
In Europe, however, such a practice would face serious legal hurdles. Valerio De Stefano, a law professor at York University in Toronto and a specialist in comparative labor and technology law, believes it would likely violate the General Data Protection Regulation (GDPR).
In Italy, the use of electronic surveillance to monitor employee productivity is explicitly prohibited. In Germany, jurisprudence allows the recording of keystrokes only in cases of serious suspected criminal activity.
Meta has so far deployed the MCI only in the United States. But as the group restructures its organization around AI, questions about expansion and limits are sure to arise.
