Meta uses employee activity data to train AI through the Model Capability Initiative

Reuters reports that Mark Zuckerberg is intensifying a full‑court push into artificial intelligence, with Meta’s employees in the United States not only building the technology but also serving as input data. Specifically, Meta is installing software that tracks employees’ activity directly on their work computers to record every mouse click, cursor movement, and keystroke. When the boss scrutinizes every click The project is named Model Capability Initiative (MCI), disclosed in an internal memo to the Meta Superintelligence Labs team. The software will run covertly on a list of applications and websites used for work. Not only does it track behavior, MCI also periodically captures screenshots of the content employees handle. Meta’s goal is clear: AI today is proficient at processing raw data but is clumsy at mimicking basic computer operations. Tasks as simple as navigating dropdown menus or using keyboard shortcuts remain challenging for large language models. By recording how thousands of employees perform daily tasks, Meta hopes to train AI agents capable of performing office work in place of humans. A Meta representative, Andy Stone, told Fortune that if you want tools to help people complete daily tasks, the model needs the most realistic examples of how people interact with computers. The internal memo also gently encourages employees that they can contribute to the company’s development simply by working as usual. Meta’s Chief Technology Officer, Andrew Bosworth, has even outlined a longer-term vision under the name Agent Transformation Accelerator (ATA). He asserts that in the future, AI agents will perform the core work, while humans shift to guiding, reviewing, and helping them improve. To realize this, Meta has implemented aggressive structural changes. The company is dissolving the boundaries between traditional job titles and replacing them with a single label: “AI builder.” Moreover, management is urging employees to use AI to write code, even if that slows progress in the short term, with the aim of generating real data for the system. This push occurs as Meta competes in a race with OpenAI and Anthropic. When internet text data becomes scarce or faces legal hurdles, exploiting behavior data from within the company becomes a new gold mine. Last year Meta spent more than $14 billion to acquire 49% of Scale AI, a data-labeling company, signaling thorough infrastructure preparation for this campaign. The specter of layoffs However, Fortune notes that this campaign could raise privacy concerns. Ifeoma Ajunwa, a law professor at Yale University, says keystroke logging has raised office surveillance to a new level, comparable to how gig economy drivers or temporary workers are managed. Although Meta pledges that the data will not be used to evaluate individual performance and that safeguards exist for sensitive content, the boundary is inherently thin. In the United States, federal law currently has no specific limits on employee monitoring, in contrast to stringent rules in Europe or Germany, where this practice can be illegal without a compelling reason. Perhaps most striking is the timing. Meta is preparing for a wave of layoffs, potentially up to 20% of the workforce, starting in May. Requiring employees to provide behavioral data to train the very tools that could replace their jobs creates a bitter paradox. This is not just a technology story but a broader question of balancing productivity and personal rights in the AI era. When Mark Zuckerberg commits to roughly $135 billion for AI infrastructure by 2026, the message is clear: those who stay at Meta must not only work with AI but become fuel to feed it until these digital agents mature enough to run the keyboard themselves. Sources: Reuters, Fortune