Meta pauses employee tracker for AI training amid privacy concerns

Meta has suspended an internal program that monitored employee keystrokes and screen content for AI training, following a staff petition and warnings that sensitive data was broadly accessible inside the company [1]. The tool, called the Model Capability Initiative, was designed to capture mouse clicks, keystrokes, and on-screen material from corporate laptops to help train Meta’s artificial intelligence models [1]. More than 1,600 workers signed a petition demanding the company stop harvesting “employee ‘computer use’ data,” arguing that “collecting and repurposing this kind of data raises serious concerns around privacy, consent, and trust in the workplace” [1]. Wired reported that data tables containing “full prompts and transcriptions, private conversations, people and performance data” were accessible to anyone inside the company, citing an internal security notice [1]. Meta confirmed the pause. “We have carefully designed this program with privacy safeguards and while we have no indication at this time that any data was improperly accessed by Meta employees, we’re pausing it while we investigate,” the company said in a statement [1]. Chief Executive Mark Zuckerberg had earlier told employees that AI models learn from “watching really smart people do things” and that “the average intelligence of the people who are at this company is significantly higher than the average set of people that you can get to do tasks” [1]. The push to harvest internal data arrives as Meta commits up to $145bn in capital expenditure this year, much of it directed at AI infrastructure [1]. The episode highlights a recurring tension inside AI firms: the drive to gather high-quality training data can collide with privacy safeguards and workforce trust. AI alignment research underscores that designers often rely on proxy goals — such as maximizing human approval — which can overlook necessary constraints or reward systems for merely appearing aligned [4]. When internal monitoring tools expose private conversations or performance data, the gap between a company’s stated safety principles and its data-collection practices widens. Broader industry dynamics have sharpened scrutiny of such initiatives. OpenAI, which catalyzed the generative-AI boom with ChatGPT in late 2022, saw roughly half of its then-employed AI safety researchers leave in 2024, citing a deprioritization of safety goals [2]. Separately, the second Trump administration’s Department of Government Efficiency — an initiative that promoted AI tools across federal agencies — drew legal rebukes after gaining widespread access to classified government data and exfiltrating it to private databases [3]. Those cases, while distinct from Meta’s internal program, illustrate how AI-linked data-access projects can trigger institutional backlash and judicial intervention. The Meta petition and the Wired disclosure suggest that even a company with Meta’s resources can stumble when it treats employee-generated data as a raw material for model training without fully locking down access controls. The company said it had no indication that data was improperly accessed by its own staff, but the investigation is ongoing [1].