Meta Trained AI on How Employees Work. Then Cut 8,000 of Them. Here's What to Invest In Instead.

The employees Meta tracked weren't doing anything wrong. They were being good workers. The tracking software deployed across Gmail, GChat, VSCode, and internal tools on April 21 was, by Zuckerberg's own description, designed to "learn from watching really smart people do things." Three weeks later, Meta announced 8,000 layoffs. The lesson here is not about surveillance. It's about what AI could observe from those workflows — and what it couldn't — and what that tells you about where to put your effort.

What happened and what the data says

Meta's tracking covered the sequence of steps employees took, the tools they used, and the outputs they produced (The Neuron, May 21). It was explicitly framed as model training. The timing — software deployed April 21, layoffs announced May 20 — was not coincidental.

In the same week, a Gartner survey of 350 executives found that companies using AI for "people amplification" outperformed those using AI to cut headcount. Of 23 S&P 500 companies that announced AI-driven layoffs, 56% saw stock declines averaging 25% — including Nike (-35%), Salesforce (-32%), and Fiverr (-54%) (The Neuron, May 18). The market is not rewarding replacement strategies.

What AI cannot reconstruct from watching you work

Josh Comeau published evidence (Hacker News, 337 points, May 22) that AI functions as a skill multiplier. An experienced developer tripled output using AI. A developer with no underlying foundation spent three hours prompting before fixing a bug manually in 30 seconds. Comeau's metaphor: AI is Iron Man's suit — powerful in the right hands, an empty shell otherwise.

The implication for the Meta story: what could a model learn from watching an employee? The steps they took, the format of their outputs, the tools they opened in sequence. What it cannot learn from a Gmail thread or a VSCode log:

• Which option was removed from the proposal before it was sent — and why
• Which stakeholder needed a call before a decision was announced
• Which metric to frame cautiously this quarter because of who's in the room

These are judgment calls that require unstated context. They exist in the professional's head, not in the record of their behavior.

What this looks like in a specific role

A procurement manager at a 350-person consumer goods company follows a workflow that looks observable from the outside: receive quotes, compare in a spreadsheet, recommend a vendor, get sign-off. Every step is logged and increasingly automatable.

What AI cannot learn from watching her inbox: that one supplier is running on thin margins and will accept a longer payment term in exchange for a contract extension. That the CFO's objection to a particular vendor isn't about price — it's about a relationship dispute from two years ago. That the right time to escalate a procurement issue is never the week before a board meeting.

The model trained on her email metadata will learn her output format. It will not learn her judgment.

A different role, the same gap

A communications director at a 90-person B2B SaaS produces a predictable set of outputs: board update slides, customer case studies, investor communications. AI can observe the format, approximate the length, and learn the structure.

What it cannot observe: which metric the lead investor is most anxious about this quarter. Whether the customer featured in the case study is in contract renewal — and should not be named until the deal closes. Which version of the board update will survive the CEO's edit and which will not.

Reading a room is not in the inbox. It builds from repeated exposure to people under pressure. It is also the skill that makes every other output more valuable.

The one thing worth taking from this

AI trained on your work can reproduce the observable pattern of what you do. It cannot reproduce why you made the calls you made — especially when the reasons were never written down. Those reasons are the investment worth making.