Meta 強制在員工電腦安裝監控，用來訓練 AI 模型，CTO：不接受拒絕

Meta is mandating the installation of keyboard and mouse tracking software on its U.S. employees' computers to train AI agent models, with company leadership explicitly stating that there is no opt-out option. (Context: Meta strongly refutes: Absolutely no cooperation with China to censor Taiwan-related content or share user data) (Background: UC study on "AI brain fog": 14% of office workers driven crazy by Agents and automation, with turnover intent 40% higher) Meta has been one of the loudest supporters of the global AI open-source movement; however, within the same office, it is doing the exact opposite: mandating that employees hand over every keystroke and every mouse movement to the company's AI system as training material. And there is no way to refuse. This happened for real this week. According to an internal announcement obtained by Business Insider, Meta has begun deploying a tracking tool called MCI (Model Capability Initiative) on the computers of its U.S. employees. The operational logic of MCI is not complex. Meta has been actively developing "AI agents" in recent years that can autonomously browse the web, fill out forms, and operate applications. However, a current shortcoming of such AI is that they do not understand how humans actually use computers. The rhythm of pressing shortcuts, the way of selecting items from dropdown menus, and the habit of switching windows between Gmail and VSCode—these "daily human operational patterns" remain a scarce piece in training data. Meta's solution is: let employees fill this gap with their daily work behaviors. The internal announcement states: "To enable AI agents to understand how humans complete daily computer tasks, we need to train models using real-world examples. This is where every Meta employee can contribute, simply by continuing to do your normal work." The data recorded by MCI includes: mouse movement trajectories, click locations, keyboard input, and screen captures. The tracking scope is limited to pre-approved work applications, including Gmail, GChat, Meta's internal AI assistant Metamate, and the code editor VSCode. Mobile phones are not within the scope of tracking. Meta states that the data has privacy protection mechanisms and will not be used for performance evaluations or other purposes. After Meta's announcement, many employees commented: "This makes me very uncomfortable. How do we opt out?" CTO Andrew Bosworth personally replied: There is "no opt-out option" for company-provided laptops. Meta's official stance is that employees are already being monitored when using company devices, and this agreement was clearly communicated upon onboarding; MCI is merely an extension of existing policies, not a brand-new one. From a legal perspective, this explanation likely holds water. But there is a significant distance between being "legal" and making employees feel respected. A report by CNBC further points out that the tracking scope of MCI actually extends to third-party websites such as Google, LinkedIn, and Wikipedia, covering a wide range of browsing behaviors during work hours. Employee concerns have also become concrete: password inputs, confidential details of new products, and even personal health conditions or immigration status could potentially be recorded during "screen captures." Meta claims that privacy protection mechanisms can filter out sensitive content, but the company has not publicly explained how this mechanism specifically operates. There is another question that the announcement did not clarify: Is the data used to train these models training "tools to help employees work," or is it training "systems to replace employees in the future"? From a technical logic standpoint, the two are not contradictory. The design goal of AI agents is precisely to learn how humans operate and then execute those operations automatically. You teach it how to filter emails in Gmail, how to switch functions in VSCode... this knowledge is universal, and whether it is used to assist you or replace you depends on how the company decides at some point in the future.