Meta tracked employee keystrokes

Meta is installing software on U.S. employees’ work computers to record keystrokes, mouse movements and clicks for artificial intelligence training. (reuters.com) The tool is called Model Capability Initiative, or MCI, and internal memos reviewed by Reuters said it will run in work-related apps and websites used by U.S.-based employees and contractors. Reuters reported the goal is to build AI agents that can perform work tasks autonomously. (reuters.com) CNBC reported on April 22 that Meta plans to capture activity on Google, LinkedIn and Wikipedia, along with hundreds of other sites and apps, using data from staffers’ actions on company machines. The Verge said the system can also capture screenshots, turning routine office work into training examples for software agents. (cnbc.com, (theverge.com) AI agents are programs built to carry out multi-step computer tasks, such as searching, clicking through menus and filling in forms, instead of only answering questions in a chat box. Training them on screen recordings and input data gives companies a way to teach those systems how people actually navigate software, not just how they write about it. (theverge.com), (reuters.com) That approach marks a shift from training mostly on text, code and labeled datasets to training on observed workflows inside office software. CNBC reported that Meta’s program extends beyond curated corpora and into live employee behavior on workplace tools and public websites. (cnbc.com), (theverge.com) Reuters reported that some employees objected in internal discussions and asked whether they could opt out. Meta told Reuters that activity on company-issued devices has long been subject to monitoring and that the new system is focused on improving AI products. (reuters.com) The program lands as Meta has accelerated its artificial intelligence push across products and infrastructure, including a broader effort to build systems that can complete more workplace actions on a user’s behalf. Using internal employee behavior as training data gives the company a source of labeled examples from people already doing those tasks every day. (reuters.com), (cnbc.com) The immediate question is not whether Meta can collect more workplace data on its own machines, but how far companies will go in turning ordinary employee computer use into raw material for AI systems. At Meta, that experiment is already moving from internal memo to installed software. (cnbc.com), (reuters.com)