Meta is implementing a controversial employee monitoring program to train its artificial intelligence models, allowing the company to capture keystrokes and mouse clicks from internal work computers across hundreds of third-party applications. The initiative, known as the Model Capability Initiative (MCI), tracks activity on widely used platforms including Google, LinkedIn, Wikipedia, Microsoft GitHub, Salesforce Slack, and Atlassian. Internal documents viewed by news outlets reveal that the tracking tool was originally intended to monitor competitor AI applications such as ChatGPT and Claude before those were removed from the list. The project aligns with Meta CEO Mark Zuckerberg's aggressive strategy to close the generative AI gap with rivals like OpenAI and Google. To accelerate progress, Zuckerberg recently hired Scale AI's Alexandr Wang to lead the Meta Superintelligence Labs. This unit recently unveiled Muse Spark, the first major model in the new Muse series designed for AI agents capable of performing complex office and coding tasks. According to a Meta spokesperson, these AI agents require real-world data on how humans interact with computers, including specific movements and button clicks, to function effectively. The rollout of the MCI tool has sparked significant internal backlash. Multiple employees have described the surveillance as dystopian, expressing fears that the system could inadvertently expose sensitive information such as user passwords, proprietary product development details, and personal data regarding health, immigration status, or family members. In an internal memo addressed to staff, a representative from the Superintelligence Labs acknowledged the need for a large, unbiased dataset to teach models how to navigate software interfaces. The memo stated that capturing on-screen content is essential for providing context to the AI. To mitigate privacy risks, the company outlined several safeguards in its communications. Officials confirmed that the tool only captures the visual output of the employee's screen as they see it and does not directly read files or attachments. The company asserts that any incidental personal information appearing in emails or other documents via the screen will not be learned by the model due to technical mitigations. Furthermore, the memo instructed employees concerned about the monitoring to refrain from conducting personal work on corporate devices to control what appears on their screens. Despite these assurances, the internal conversation on chat boards remained heated, with staff debating the ethical implications of such extensive data collection. While Meta has confirmed the existence of the project, a spokesperson declined to comment further on the specific list of tracked sites, maintaining that the data collected serves solely to train AI agents and will not be used for any other purpose. As Meta continues to expand its AI capabilities, this initiative highlights the ongoing tension between technological advancement and employee privacy within the tech industry.