Meta Platforms announced on April 22, 2026, the launch of a new internal data collection program titled the Model Capability Initiative. This program is designed to accelerate the development of the company’s next generation of artificial intelligence by directly observing and recording the digital workflows of its United States-based workforce. Under the initiative, Meta is deploying specialized tracking software across corporate-issued computers to capture granular behavioral data, including mouse movements, click patterns, keystrokes, and periodic screenshots.
The primary objective of the Model Capability Initiative is to provide high-fidelity training data for Meta’s autonomous AI agents. By analyzing how human employees navigate complex software environments, manage communications, and execute technical tasks, the company aims to refine the reasoning and execution capabilities of its Large Language Models. Meta executives stated that the data will be used to train agents capable of performing end-to-end workplace functions, ranging from software debugging to administrative scheduling and project management.
According to internal documentation released alongside the announcement, the tracking software operates in the background of macOS and Windows environments used by Meta staff. The system records the sequence of actions taken to complete specific tickets or projects, creating a map of human decision-making processes. Meta’s Chief Product Officer, Chris Cox, noted in a memo to staff that this behavioral cloning approach is essential for bridging the gap between static text-based training and the dynamic requirements of real-world digital labor.
The initiative initially applies to thousands of employees across various departments, including engineering, product management, and legal operations. Meta has confirmed that the data collection is mandatory for US-based personnel using company hardware, though the company has established a dedicated Model Privacy Task Force to oversee the anonymization of sensitive personal information captured during the process. The company clarified that the screenshots and keystroke logs would be processed by automated systems to strip out passwords and private financial data before being ingested into the training pipeline.
This move follows Meta’s broader strategy to pivot toward agentic AI, a shift highlighted during the company’s most recent quarterly earnings reports. By utilizing its own workforce as a live laboratory, Meta intends to reduce its reliance on third-party datasets and public web scraping, which have faced increasing regulatory and legal scrutiny. The company has not yet disclosed a timeline for when the resulting AI agents will be integrated into its commercial product offerings or internal infrastructure.