Meta’s Employee Backlash Shows the Office Is Becoming AI’s New Data Mine

The unsettling part of Meta’s employee tracker is the intimacy of what it tried to turn into data. Meta paused its Model Capability Initiative after more than 1,600 employees signed a petition against it, The Guardian recently reported. The tool had reportedly tracked keystrokes, mouse clicks, and screen content from corporate laptops to collect data for AI training.

For years, companies have treated workplace monitoring as a security or productivity issue to protect the network, prevent leaks, track compliance and measure output. Employees may dislike parts of that bargain, but the logic is familiar.

AI changes the bargain because the company is no longer only looking at what someone produces. It is looking at how they get there. The real machinery of knowledge work is far more personal than the finished output. Inside a company like Meta, thousands of highly skilled people are solving hard problems all day, in context, using tools, judgement, memory, and expertise.

Company CEO Mark Zuckerberg reportedly told employees that AI models learn from “watching really smart people do things.” As a technical thesis, it makes sense. A model that observes strong engineers, researchers, and product teams may learn better patterns than a model trained only on public material.

As a workplace proposition, it lands very differently. People do not experience their laptops as training environments. They experience it as the place where they draft badly before drafting well, complain privately before replying professionally, test ideas before defending them, and make mistakes before correcting them. The screen is where work becomes human before it becomes corporate.

Meta employees were objecting to more than observation. They were objecting to the conversion of working behavior into model fuel. The Guardian cited Wired’s reporting that MCI data collected from corporate laptops had been accessible inside the company, including prompts, transcriptions, private conversations, people data, and performance related material. Meta said it had no indication that data had been improperly accessed and paused the programme while investigating.

Once workplace activity becomes a dataset, it can travel. It can be searched, retained, linked, inferred from, and interpreted outside the moment in which it was created. Employees need to know what is captured, what is excluded, who can see it, how long it lives, and whether it trains broader systems. In a workplace, consent is complicated because the power relationship is uneven. People can agree to a policy while still feeling they had little real choice.

There is also a practical problem for companies that want to train AI on human work. Surveillance changes the behaviour it measures. When people feel watched, the work becomes more guarded. Sensitive conversations move elsewhere, rough thinking disappears from shared tools, and the workplace starts to look cleaner than it really is.

That is the paradox. AI learns best from the messy parts of expertise, judgement, correction, doubt, and revision, yet those are the first things people protect when monitoring feels intrusive.

The boundary is simple enough to state but difficult to govern: AI at work should help people use their expertise, rather than quietly absorb it.

Meta’s tracker may have stopped for now but the issue will grow more and more relevant in the next few years. Which begs the question: how much of a person’s working mind should become corporate training material?