The move, disclosed in an internal memo seen by Reuters, is framed as a way to teach AI agents how humans navigate software. Critics say it is workplace surveillance under a different name.
Meta is installing new tracking software on US-based employees’ work computers to capture mouse movements, clicks, keystrokes, and occasional screenshots, which will be fed into the company’s AI model training pipeline, Reuters has reported.
The tool, called the Model Capability Initiative, was disclosed to staff this week in a memo distributed via a channel belonging to the Meta Superintelligence Labs team.
According to Reuters, which saw the memo, it will run on a designated list of work apps and websites. Staff were told they can “do their part to help by just doing their daily work.”
Meta’s stated rationale is that building agents capable of navigating software on behalf of users requires training data drawn from actual human computer use, specifically the micro-behaviours, such as navigating dropdown menus and using keyboard shortcuts, that AI models struggle to replicate from general web data alone.
“If we’re building agents to help people complete everyday tasks using computers, our models need real examples of how people actually use them, things like mouse movements, clicking buttons, and navigating dropdown menus,” a Meta spokesperson said.
The company added that safeguards are in place to protect sensitive content and that the data will not be used for any purpose other than model training.
The memo reportedly framed the effort as part of Meta’s “AI for Work” programme, which has been renamed the Agent Transformation Accelerator.
The Superintelligence Labs division, which leads Meta’s most ambitious AI research, is now led by Alexandr Wang, the former CEO of Scale AI, the data-labelling firm in which Meta acquired a 49% stake for more than $14 billion last year.
The connection is significant: Scale AI’s core business is producing training data at industrial scale, and Wang’s appointment signals that Meta is treating data collection as a core strategic competency rather than a procurement exercise.
The move puts a blunt question about the labour relationship between AI companies and their employees into public view. For Meta workers, the dynamic is unusually direct: they are, in a literal sense, being asked to generate the training data that will teach AI agents to replicate their own computer-use behaviour.
Meta is not the first company to mine internal workflows for AI training, in January, OpenAI was reported to be asking contractors to upload samples of real work products, but the combination of keystroke capture and screenshot collection represents a more systemic and automated approach.
The company’s assurance that the data will not be used for performance monitoring has been met with scepticism in public commentary, though no evidence has emerged that it is doing so.
The tracking software will run only on a designated list of apps and websites, not across all computer activity. Meta has not disclosed which applications are included.
The data collection appears to be limited to US employees at this stage; EU and UK data protection law would require explicit legal bases and potentially explicit consent for equivalent measures in those jurisdictions.
