Meta will record employees’ keystrokes and use it to train its AI models

The Data Behind the Agents

In its ongoing quest for high-quality training data, Meta has turned its focus inward, confirming plans to record employee keystrokes and mouse movements to train its next generation of AI models. The initiative, first reported by Reuters, underscores the intense industry-wide pressure to find novel data sources that can teach AI to perform complex, human-like tasks on a computer, effectively turning a portion of its workforce's daily activity into a proprietary data asset.

According to a statement from Meta, the data is being collected via a new internal tool designed to help build AI agents. The company argues that to create models that can assist people with computer-based tasks, they require real-world examples of user interaction. The data collected will not be used for any other purpose, and the company states that safeguards are in place to protect sensitive information.

• Data Captured: Mouse movements, keystrokes, button clicks, and dropdown menu navigation.
• Stated Purpose: To train AI agents on how people use computer applications.
• Safeguards: Internal controls are in place to protect sensitive content.
• Scope: The tool will capture inputs on “certain applications” internally.

This move by Meta is part of a larger, and increasingly troublesome, trend in the AI industry where the definition of training data is rapidly expanding. It follows recent reports of startups being acquired solely for their internal corporate communications, such as Slack archives and Jira tickets, to be repurposed for model training. This raises significant new questions about employee privacy and the ethical boundaries of repurposing work-related data for AI development, potentially setting a new standard for how tech companies leverage internal resources.

While public and licensed data have powered foundational models, the next frontier for AI agents requires fine-grained, procedural data. Meta's internal collection strategy is a direct attempt to build a proprietary dataset of human-computer interaction, bypassing the noise of the open web to create more capable and reliable agents.