The ‘Model Capability Initiative’ logs mouse movements, keystrokes, and screenshots for AI training purposes.
Meta has begun using employee activities to train its AI systems. According to Reuters, the company is implementing a tool known as Model Capability Initiative (MCI) on the computers of its US-based workers. This tool runs within work-related applications and websites, capturing mouse movements, clicks, keystrokes, and occasional screen captures.
The data collected will train Meta’s AI models to better mimic human interaction with computers, thereby enhancing task automation—tasks similar to those carried out by Meta employees. Reuters notes that MCI’s data won’t be used for performance evaluations.
Tracy Clayton, a Meta spokesperson, stated to The Verge, “To develop agents that aid with everyday computer tasks, our models need real-life examples—mouse movements, button clicks, navigating dropdown menus. We’re deploying an internal tool for data capture on certain apps, safeguarding sensitive content, and ensuring the data’s sole purpose is model training.”
In an internal memo, Meta CTO Andrew Bosworth informed employees about expanding data collection initiatives, including MCI, for the Agent Transformation Accelerator (ATA). Bosworth highlighted that the goal is to have agents primarily execute tasks, with humans overseeing and guiding improvements.