Tech companies are getting desperate for AI training data, and Meta’s latest move shows just how far they’re willing to go. According to Reuters reporting, the company is planning to use data harvested directly from its own employees’ computer usage, including mouse movements, keystrokes, and button clicks, to train its AI models.
This isn’t some dystopian fiction plot point. It’s happening now, and it reveals something troubling about how the AI industry operates.
The Data Diet Gets Personal
Meta’s approach is straightforward in its pitch but unsettling in practice. When reached for comment, a Meta spokesperson explained that building AI agents to help people complete everyday tasks requires “real examples of how people actually use computers.” The company says it’s launching an internal tool to capture these inputs on certain applications, with safeguards in place to protect sensitive content.
That’s the official framing, anyway. What’s really happening is that Meta is treating its own workforce as a living laboratory, converting the mundane activities of everyday work into fuel for its AI ambitions. Your scroll. Your click. Your keystroke. All potentially grist for the algorithm mill.
The company insists the data won’t be used for other purposes and that safeguards exist to filter out sensitive information. Whether those safeguards hold up in practice, or whether they’re robust enough to actually protect the privacy of 67,000-plus employees, remains an open question.
A Symptom of the Larger Problem
This Meta initiative isn’t an isolated incident. It’s part of a broader trend in technology that shows companies are rapidly running out of obvious places to source training data. Last week, it emerged that dead startups were being scavenged for their internal communications, with Slack archives, Jira tickets, and other corporate messaging platforms being converted into AI training material.
Think about that for a moment. The casual conversations, the half-baked ideas, the frustrated venting that happens on company Slack channels, the technical documentation on internal wikis, the decision-making threads buried in old projects. All of it potentially valuable to someone building an AI model.
The underlying issue is simple: AI models are voracious consumers of data. The better the data, the better the model. And the tech industry has already scraped most of the easily accessible corners of the internet. So they’re now turning inward, looking at the data sitting in corporate repositories, employee devices, and archived communications.
The Privacy Reckoning That Hasn’t Arrived
What’s remarkable is how little pushback this generates. Meta’s announcement came with a statement emphasizing safeguards, and the story largely landed as a factual report rather than a scandal. But if a company announced it was monitoring employee mouse movements and keystrokes for any other purpose, there’d be immediate outcry about surveillance and worker privacy.
Rebrand it as “AI training,” and suddenly it becomes a routine business decision.
The problem isn’t just Meta. It’s the entire incentive structure. AI models need data. Companies have data. The data sitting in your inbox, your Slack messages, your browsing history inside corporate systems is already there. The marginal cost of using it is nearly zero. The compliance risk, at least so far, has been minimal. Why wouldn’t companies harvest it?
The real question is whether employees, regulators, or privacy advocates will eventually decide this crosses a line. Because right now, the answer seems to be that they won’t, or at least not quickly enough to stop the practice from becoming routine.
