Meta is asking U.S.-based employees to generate AI training data the old-fashioned way: by doing their jobs. New internal software will record mouse movements, clicks, and keystrokes on selected work apps and websites, while also taking occasional screen snapshots, so the company can teach its models how people actually use computers.
Meta wants its future AI agents to handle everyday office tasks, and that means learning the unglamorous bits like dropdown menus, keyboard shortcuts, and basic navigation. In other words, the bots are being trained on the same routine computer work that already eats everyone’s day. Efficient? Sure. Slightly unsettling? Also yes.
What Meta says the tracking will capture
According to an internal memo shared in Meta’s model-building team channel, the tool will focus on work-related apps and websites rather than roaming across the whole machine. The snapshots are meant to add context, helping the system understand what a user was doing when a click or keystroke happened.
Meta spokesperson Andy Stone said the data will be used only for model training, not performance reviews. That reassurance matters, because employee monitoring software has a long and dreary history of creeping from ”productivity” into ”pressure.” Meta is trying to draw a bright line before anyone starts imagining the boss’s dashboard.
- Mouse movements
- Clicks
- Keystrokes
- Occasional screen snapshots
Why Meta wants real employee behavior
The company’s bet is that artificial agents need more than clean demos and synthetic examples. If a model is going to operate inside business software, it has to learn how humans stumble through it in real life, not how they behave in a polished lab test. That is a smart move for model quality, and a pretty direct reminder that AI still struggles with annoyingly ordinary tasks.
Meta is also under pressure to show that its AI work is more than hype around chatbots and glasses. Competitors across the industry have been racing to build agents that can execute tasks rather than just answer questions, and the winning models will need training data that reflects messy, repetitive office behavior. This is where Meta is trying to turn its own workforce into a data engine.
The employee privacy trade-off
The awkward part is obvious: even if the system is scoped to work tools, it is still watching how people use their computers. Meta says safeguards are in place to protect sensitive content, but the company is asking workers to trust that metadata, screenshots, and keystroke trails will stay safely inside a model-training pipeline. That’s a familiar promise in tech, and employees have heard some version of it before.
Still, the move fits a broader pattern in AI development. The most valuable training data is increasingly the stuff that looks mundane to humans but is rich with signal for models: decisions, sequences, mistakes, and corrections. If Meta can pull this off without spooking its own staff, it will have found a cheap and continuous source of examples. If not, it may learn that the hardest system to train is the one sitting at the next desk.