Meta MCI: Why Meta Is Training AI on Its Employees

In April, Meta started recording its US employees' mouse movements, button clicks, keystrokes, and screenshots from designated work apps and websites. The data is feeding the Model Capability Initiative, an internal program to train AI agents on how white-collar work actually gets done. On April 30, at an internal all-hands, Zuckerberg defended the program. The audio surfaced publicly on May 19.

Read that sentence stripped of the corporate framing. What it says is: my engineers are higher-signal than the people I could hire to generate training data, so I'm going to record what my engineers do instead.

The same day the audio surfaced, Meta issued layoff notifications to approximately 8,000 of those same employees.

What is Meta's Model Capability Initiative?

MCI captures mouse movements, button clicks, keystrokes, and screenshots from designated work apps and websites, per TechCrunch's April 21 disclosure. Scope includes US rank-and-file employees on work computers. The tracked surfaces include external sites like Google, LinkedIn, and Wikipedia along with internal enterprise tools, per CNBC. Meta's stated rationale, from the program's accompanying spokesperson statement, is operational. The company wants AI agents that can complete everyday tasks using computers. The agents need real examples of how people use them. The fastest path to those examples runs through the people already using the computers Meta owns.

Two weeks later, Meta announced a roughly 10% workforce reduction inside its $135 billion 2026 AI capex commitment. The layoff notifications went out May 19, the same day the leaked Zuckerberg audio surfaced publicly. The timeline does not need editorializing. The same human capital being captured as training data is being eliminated as a cost line. The two moves are not unrelated. They are pieces of the same strategy.

Is this just a privacy story?

The public framing so far treats MCI as a privacy and labor-rights story. That framing is correct. The data capture is invasive. Employees are right to object. The 1,000+ signatures on the internal petition reflect the appropriate institutional pushback. Workers calling the program dystopian are using exactly the right word. The legal exposure under US labor law, EU data-protection rules, and ongoing state-level privacy regulation is real and serious.

The privacy reading isn't wrong. It's incomplete.

The privacy framing treats MCI as a deviation from normal corporate practice. It isn't. MCI is consistent with a specific choice Meta has made about what its workforce is for. The deeper question is what kind of company turns its own engineers into a training corpus, lays them off, and books both moves as AI strategy. The privacy debate is the visible surface of that choice. The choice itself is the news.

In the same 30 days Meta deployed MCI, three other frontier labs made the opposite choice.

Why are other AI labs doing the opposite?

I wrote two days ago about the lab deployment pivot. OpenAI launched the OpenAI Deployment Company with $4 billion in initial capital at a $10 billion valuation, anchored by the acquisition of Tomoro and its 150 Forward-Deployed Engineers. Anthropic closed a $1.5 billion deployment joint venture with Blackstone, Hellman & Friedman, Goldman Sachs, and others. Google committed $750 million to existing consulting firms operating on the same embedded model. Combined: $6.25 billion in capital allocation in 30 days, all flowing toward the same structural conclusion.

The Anthropic JV announcement described how engagements would work:

An engagement might begin with the company's engineering team sitting down with clinicians and IT staff to build tools that fit into the workflows that staff already use. Engagements like this will run across mid-sized companies across industries, each shaped by the people closest to the work.

Shaped by the people closest to the work. That phrase is a near-verbatim restatement of our manifesto pillar closest-to-the-problem, published in mid-May, before Anthropic's announcement.

Same underlying question facing both Meta and the FDE labs. Where does the workflow context that makes AI agents useful actually come from? Two completely different answers. Meta's answer: capture it from the people who already produce it, then keep what you captured when you no longer need the people. The FDE labs' answer: pay engineers to embed inside customer organizations and produce the context as part of the engagement.

The contrast pair is the news. The privacy debate sits on top of it.

What does the extraction model actually look like?

The two models in May 2026 are fully visible.

Extraction model (Meta MCI). Your engineers are the highest-quality available training data. Surveil them. Capture their patterns. Train your AI on what they produce. When the AI is good enough, lay off the engineers. Their patterns are now in the model.

Collaboration model (Anthropic, OpenAI, Google deployment arms). Your engineers are collaborators whose value emerges in engagement with the customer's domain experts. Pay them to embed inside client organizations. Co-create the build alongside the people closest to the work. The deliverable is the engagement itself, not the captured trace of it.

The two models share an input. Both assume that AI agents need real workflow context to ship useful outcomes. The agreement ends there. The choice about how to source that context is the structural fork. Extraction sources it from people who are about to become unnecessary. Collaboration sources it from people whose presence in the work is the deliverable.

This is not a values debate dressed up as strategy. It is a strategy debate with values consequences. The extraction model assumes that workflow patterns are the bottleneck and humans are the workaround until you have enough patterns. The collaboration model assumes that workflow patterns alone are insufficient because the build depends on the domain expert being in the engagement, shaping decisions that no pattern can replicate.

How do you read a vendor's intent on AI labor?

Watch the verb.

"Our models need real examples of how people actually use them" puts the model at the center; the human is the source. "Engagements shaped by the people closest to the work" puts the human at the center; the engagement is the deliverable. Vendor language is rarely accidental. Companies that intend to extract speak in extractive verbs even when they are being careful. Companies that intend to collaborate speak in collaborative verbs even when they are being economically aggressive.

If you are evaluating an AI vendor's relationship with humans, theirs or yours, the verb is the tell.

The historical pillar this all extends is who-creates-is-changing. Every prior wave in computing (GUIs, CMSs, the web, mobile) lowered a barrier and brought new people into the work. The AI wave is bringing domain experts into the build at scale. That's the optimistic version of the shift. The dark version is what MCI looks like: a wave that brought domain experts in as data sources, then dropped them once their patterns were captured. Both versions of "who creates is changing" are running concurrently in May 2026. The technology is the same in both. The choice is structural.

What does this mean for builders?

If you are building a company that depends on humans (yours, your customers', your domain experts'), MCI versus FDE is the choice you are making every quarter whether you name it or not.

The default for AI-augmented organizations right now is closer to MCI than to FDE. Extract patterns. Automate roles. Treat the workforce as an interim training corpus until the model is good enough. The active alternative is the one the FDE labs just put $6.25 billion behind and the one our recent posts have been describing: humans as collaborators, domain experts in the build, engagements shaped by the people closest to the problem.

Both choices are visible. Neither is forced by the technology.

Meta has made a coherent strategic choice. Zuckerberg is correct on his own terms: his engineers are higher-signal than crowd workers, and capturing their patterns is the rational move inside the extraction model. The other labs have made an equally coherent but opposite choice. The deeper question isn't who is right about model architecture or labor economics. The question is what kind of company you intend to build when you decide whether the humans you depend on are collaborators or training data.

The privacy debate is real. The labor objection is right. They sit on top of the deeper question. The model you choose for the human and AI relationship is the model you become.