Meta tracks employee activity

Meta is logging how some employees use their work computers, including keystrokes and mouse clicks, to build better artificial intelligence systems. (cnbc.com) Reuters reported on April 21 that the new tool, called Model Capability Initiative, is being installed on U.S.-based employees’ machines and can capture activity across work apps and websites, plus occasional screen snapshots. Meta told staff the data would be used to train AI agents that can carry out office tasks on computers. (news.yahoo.com) CNBC reported a day later that the monitored list includes hundreds of sites and services, among them Google, LinkedIn, Wikipedia, GitHub, Slack, Atlassian products, Threads and Manus. Internal messages reviewed by CNBC said the list was still changing and at one point also included ChatGPT and Claude. (cnbc.com) The basic idea is simple: Meta wants examples of how humans move through software, from clicking buttons to using keyboard shortcuts and dropdown menus. Meta spokesperson Andy Stone said the company needs those examples if it wants agents to help people complete everyday tasks on computers. (news.yahoo.com) The project lands in the middle of a broader rework of Meta around AI. In a memo described by Reuters, Chief Technology Officer Andrew Bosworth said the company’s “AI for Work” effort had been rebranded as Agent Transformation Accelerator, with a goal of having agents do more of the work while employees direct and review them. (news.yahoo.com) That push has accelerated since last summer, when Mark Zuckerberg began spending heavily to rebuild Meta’s AI group. CNBC reported that Meta brought in Scale AI co-founder Alexandr Wang to lead Meta Superintelligence Labs, and on April 8 the company introduced Muse Spark, its first major model from that unit. (cnbc.com) Meta says the tracking data is for model training, not employee reviews. Stone told Reuters the company would not use Model Capability Initiative data for performance assessments and said safeguards were in place for “sensitive content,” though Reuters said Meta did not explain which categories would be excluded. (news.yahoo.com) Employees still raised concerns internally after the rollout memos circulated, according to CNBC and Reuters, and Meta Superintelligence Labs staff discussed surveillance and privacy on internal chat boards. The dispute is less about whether Meta is building AI agents than about how much worker behavior the company is willing to turn into training data to do it. (cnbc.com)