Meta's Leaked Plan to Surveil Employees and Use That Data to Train AI
TL;DR
Meta’s leaked plan is explicit about using employee activity as AI training data — the memo says Meta wants to capture mouse movements, clicks, keystrokes, and occasional screenshots in work apps so agents can learn from how people actually use computers.
The hosts say the real endgame is obvious: train agents to do the work, then shrink the workforce — they point to CTO Andrew Bosworth’s line that agents will “primarily do the work” while humans direct, review, and improve.
Meta’s layoffs sit right next to an enormous AI spending surge — the company is cutting about 10% of staff starting May 20 while guiding $115 billion to $135 billion in AI infrastructure capex, nearly double 2025 spending.
The most unsettling part isn’t monitoring itself — it’s monitoring plus replacement — the hosts note employee surveillance on work devices is old news, but using that data to build computer-use agents that can replace those same workers is the new, murkier twist.
This isn’t just a Meta story; the hosts argue it’s becoming a playbook across tech — they mention xAI, a startup selling this exact capability, and Scale AI paying professionals hundreds of dollars an hour to record workflows for model training.
The talent risk is cultural, not just technical — Paul Roetzer argues great employees will start asking whether they want to work somewhere “tracking literally everything you do” while likely training your replacement.
The Breakdown
The leak: Meta wants to watch how employees work
The episode opens with a leaked Meta memo describing software that would track U.S. employees’ mouse movements, clicks, keystrokes, and occasional screenshots across approved work apps and sites. The stated goal is to teach AI agents how humans actually use computers, while Meta insists it won’t read files or attachments, use the data for performance reviews, or retain incidental personal information.
Bosworth’s quote makes the strategy hard to miss
What really sharpens the story is CTO Andrew Bosworth’s framing of the destination: agents will “primarily do the work,” and people will direct, review, and improve them. That’s why the hosts immediately ask the blunt question most employees would ask: are they training agents to do the humans’ jobs and then cutting the humans?
Layoffs and capex land in the same breath
The conversation then ties the memo to Meta’s workforce cuts: roughly 10% of staff, beginning May 20, with more expected in the second half of 2026. Mike stops short of overclaiming, but points directly at Meta’s $115 billion to $135 billion AI infrastructure capex guidance — nearly double 2025 — as the giant “other investment” likely driving some of the efficiency push.
Paul’s reality check: this has been happening already
Paul says the surveillance angle is not new at all, recalling a story from two years ago when an employee at another social network asked him if her company had been recording everything on her computer. His point is that the capability has existed for years; what’s changing is that companies are now much more open about using it to generate training data for AI systems.
The human cost: who wants to sit there watching an agent work?
Paul turns from the tech to the culture question: what kind of organization do you become when employees know they may be training their replacement? He uses a vivid image of an assembly line where your job becomes sitting there for eight hours watching an agent click around, and says that doesn’t sound like a career ambitious talent will choose.
Meta is just the clearest example, not the only one
The hosts are careful not to treat this as a uniquely Meta scandal, because companies have monitored work devices for security and compliance for years. What feels new — and “murky really quick” — is the second lane: collecting workflow data specifically to train computer-use agents that can replicate marketing, sales, customer success, HR, finance, and operations work.
The broader pattern: xAI, Scale AI, and the market for human workflows
By the end, the discussion widens beyond Meta entirely. Paul says xAI reportedly updated employee terms so work could feed Grok training, and explains that firms like Scale AI have paid lawyers and consultants as much as $500 an hour to record their desktop workflows so models can learn those jobs — prompting Mike to say this has “Alexander Wang’s fingerprints all over it.”