Meta will capture employee keystrokes and mouse data to train AI models

Meta is implementing a new internal tool designed to record employee input data—including mouse movements, keystrokes, and button clicks—to feed its AI model training pipeline. The system targets specific applications and aims to provide real-world behavioral examples that help AI agents learn how to assist users in completing everyday computer tasks.

According to a Meta spokesperson, the initiative addresses a fundamental need in agent development: access to authentic examples of how people interact with software interfaces. The company states that protective measures are embedded to filter sensitive information before data reaches training systems, and that the captured data is segregated for AI training purposes only.

The move exemplifies a widening practice across the tech industry to source training data from internal corporate operations. Previous reporting has documented efforts to repurpose archived internal communications—including Slack messages and project management tickets—as training material, underscoring the competition for fresh datasets amid the rapid expansion of AI capabilities.