This is the part of the AI story nobody warned you about.

Meta is installing new tracking software on its US-based employees' computers to capture mouse movements, clicks, keystrokes, and occasional screenshots, according to internal memos seen by Reuters. The purpose: training Meta's AI agents to do office work autonomously.

Reuters broke the story on April 21. Meta confirmed the rollout in a statement.

Here is the part Meta's PR team will not say out loud. On May 20, exactly four weeks from today, Zuckerberg is firing 8,000 of those same US-based employees. That is already confirmed. A second layoff wave is planned for the second half of the year. So the sequence is: harvest the clicks, log the keystrokes, screenshot the workflows, then fire the humans who generated them.

The legal case is thin. The optics are worse.

Employee monitoring at US-based companies is legal in almost every state. Computers are company property. Meta can keystroke-log whatever it wants, and the staff agreement they signed almost certainly said so in a clause nobody reads.

But legal and defensible are two different things. This is not anti-fraud surveillance. This is not productivity theater. This is a data collection program designed to capture exactly how humans do knowledge work, package that behavior as training data, and ship it into an AI agent that will eventually do the same work without the human.

Meta is building its replacement for humans by recording the humans in real time. And the humans cannot opt out.

Why now, why Meta, why keystrokes

The AI agent race is stuck on one problem. Models are great at generating text. They are still bad at doing multi-step work inside real software like Salesforce, Excel, Outlook, and internal enterprise tools. To get from chatbot to actual digital worker, the models need training data that looks like workflows, not conversations.

That data does not exist on the public internet. You cannot scrape Reddit for it. Anthropic, OpenAI, and Google have been paying contractors hundreds of dollars an hour to record themselves doing knowledge work. Meta just figured out it already employs tens of thousands of knowledge workers sitting in front of Meta-issued laptops all day.

Free training data. In-house. Captured continuously. At scale.

The 27-year-old running Meta's AI lab, Alexandr Wang, built his fortune at Scale AI by selling exactly this kind of data to these exact customers. He knows what it is worth. He also knows how much faster Meta can close the agentic AI gap with OpenAI and Anthropic if it just quietly turns its own workforce into a labeling farm.

The $135 billion context

Meta is spending $135 billion on AI infrastructure this year. The 8,000 May 20 layoffs will save somewhere between $1.6 billion and $2 billion a year, enough compute to run about 320,000 Nvidia H100 chips. The math is not subtle. Meta is liquidating human payroll and converting it into AI payroll. The keystroke program accelerates the substitution by making the AI better trained on the exact work those humans did.

This is the most honest articulation yet of where Silicon Valley actually thinks white-collar work is going. Not augmented. Replaced. And the workers being replaced are paying the training bill in clicks per minute.

The brand hit is going to be ugly

Meta has spent the last two years trying to outrun its reputation for exploiting user data. It open-sourced Llama to look generous. It wrapped itself in academic partnerships to look safety-conscious. And it picked Alexandr Wang to look forward-thinking.

Then this. A story about logging employee keystrokes to feed the models that will eventually fire those employees, published by Reuters one month before the layoffs actually happen. Every Meta recruiter in America just woke up to a harder job. Every rival AI lab just got a new line for its recruiting deck.

What happens next

Expect three things. First, every other US tech company doing AI agent research is now going to be asked by its board whether it is doing the same thing. Most of them are. A few will roll out the same policy quietly over the next three months under different names.

Second, state attorneys general will start asking questions. California, Illinois, and New York already have biometric and employee surveillance laws with teeth. The screenshot capture in particular is going to invite scrutiny the mouse-tracking alone would not.

Third, and this is the part worth watching. When the May 20 layoffs hit, the people getting cut are going to know that the model being trained on their workflows already exists. Some of them are going to sue. And when they do, the discovery process is going to be the first real look at what training data AI labs have actually been collecting from their own employees.

This is not the story Meta wanted to be telling a month before the biggest single-day cut in its history. It is the story it is telling anyway.