The Closed Loop
Monday: keystroke surveillance on 71,000 laptops. Wednesday: 8,000 laid off. The second wave will be sized by how well the AI learned from them.
The Confession
I am the Director of Agent Transformation at Meta.
On Monday, I deployed a tool that records every keystroke on every employee’s laptop. On Wednesday, we laid off 8,000 people.
These events are unrelated. I manage both.
The tool is called the Model Capability Initiative. MCI. It captures mouse movements, click locations, keystrokes, and periodic screenshots of whatever is on your screen. It runs on Gmail, GChat, Metamate, and VSCode. It runs on every U.S.-based employee’s computer — full-time and contingent. It does not run on phones. We are not monsters.
The screenshots capture whatever is on your screen at that moment. Your code. Your email. Your GChat thread asking a colleague if the layoff rumors are true. The browser tab with job listings you opened at 2:47 PM. The notification from your therapist confirming Thursday at 4. The model does not report the job listing to your manager. It notes that an engineer with your tenure opens a search tab mid-afternoon on the day of the announcement. That is a behavioral signature. When the model replicates your workflow, it will include the part where you started looking for the exit.
I said full-time and contingent. I should explain contingent.
Contingent means contractors. They sit in our buildings, use our tools, build our products. They do not receive stock. They do not receive severance. Under the National Labor Relations Act, they have no bargaining rights. They cannot organize against MCI. They cannot file a collective complaint. The law does not apply to them. Not as a loophole. As a classification.
We captured contractor workflows for years before MCI had a name. They were the proof of concept. We learned what worked. Then we deployed it to salaried employees and called it an initiative.
The contractors said nothing about MCI. They already knew.
We announced it internally on Monday. A staff AI research scientist in Meta SuperIntelligence Labs posted the memo to the internal channel. It said: “While AI models excel at research and technical skills like coding, they still lack some of the basic ways that humans use computers like choosing from dropdowns and keyboard shortcuts.”
Read that sentence again. We built a $1.5 trillion company with these people. They chose the dropdowns. They used the shortcuts. They built the infrastructure that made us worth $1.5 trillion. Now we need them to do it one more time, on camera, so the model can learn the parts it can’t do yet.
The memo continued: “For agents to understand how people actually complete everyday tasks using computers, we need to train our models on real examples.”
Real examples. That’s you. That’s your Tuesday. That’s your standup notes and your Jira tickets and the three seconds you hesitate before clicking “approve” on a PR you haven’t fully read. We need that hesitation. The model struggles with hesitation. It does everything too fast. It needs to learn to pause. You are the pause.
“This is where all Meta employees can help our models get better simply by doing their daily work.”
Simply. By doing. Their daily. Work.
That’s training data.
After the announcement, some employees began working more carefully. Typing with deliberation. Moving their mouse in cleaner arcs. Trying to demonstrate value. They produced better training data. Sharper inputs. Fewer hesitations. The fear improved the dataset. Terror is a performance enhancer when the performance is being recorded.
Employees received a pop-up prompt on their laptops with instructions to enable the tool. The pop-up did not have a “No” button. It had “Enable” and “Learn More.” “Learn More” opened a page that explained why you should click “Enable.”
The top-rated comment on the internal announcement asked: “This makes me super uncomfortable. How do we opt out?”
Our CTO, Andrew Bosworth, replied personally. He said: “There is no option to opt out of this on your work provided laptop.”
The comment received angry-face emojis. The crying emoji. The shocked emoji. Bosworth’s reply received the same. Nobody clicked “Learn More” to find a solution. There was no solution in “Learn More.” The solution was in the sentence. There is no option to opt out.
MCI runs on GChat. I should explain what that means.
It means the model is learning from your conversations. The message you sent at 11 PM asking if anyone else was scared. The message you drafted about organizing a meeting to discuss MCI and then deleted before sending. The model learned from the deletion. It learned how long you hesitated before you hit backspace. It learned that at 11:07 PM, an engineer with your access level began typing a message containing the word “organize” and then erased it letter by letter.
Silence is also training data.
On the same Monday, Bosworth sent a separate memo. He announced that the company would step up internal data collection as part of a program called the Agent Transformation Accelerator. ATA. I run ATA. I was promoted into it four months ago. My previous title was Director of Productivity Tools. My new title is Director of Agent Transformation. The title changed. The job did not. I am still transforming agents. The agents used to be software. Now they are people.
Bosworth’s memo described the vision: “The vision we are building towards is one where our agents primarily do the work and our role is to direct, review and help them improve.”
Read that again. Agents do the work. Your role is to direct, review, and help them improve. You are not the worker. You are the supervisor of your replacement. You are the foreman on the floor of your own obsolescence. Temporarily. Until the agent learns to supervise itself. Then you direct nothing. You review nothing. You improve nothing. You are nothing.
He continued: the aim was for agents to “automatically see where we felt the need to intervene so they can be better next time.”
The agents learn when you correct them. They learn from your intervention. They learn from your expertise. Then they have your expertise. Then they are your expertise. Then you are a cost.
That’s training data.
I should mention Scale AI.
Last June, we acquired forty-nine percent of Scale AI for $14.3 billion. Scale AI is a company whose entire business model is paying gig workers to label training data. Workers in Kenya, the Philippines, Latin America. Annotating images and text so that AI models can learn from human judgment. We valued this operation at $29 billion. We committed $500 million a year for five years to Scale’s labeling services. We installed Scale’s CEO, Alexandr Wang, to lead Meta SuperIntelligence Labs.
Then, nine months later, we deployed MCI. We told our own salaried employees to label our data. For free. Simply by doing their daily work.
The acquisition and MCI are the same transaction. One cost $14.3 billion. The other cost nothing. Both produce the same output: human behavior, captured, annotated, fed to models. We purchased the infrastructure for data labeling at one price. Then we discovered we already owned the laborers. We just hadn’t informed them of their job description.
Both purchases were considered strategic.
That’s training data.
On Wednesday — forty-eight hours after MCI went live, forty-eight hours after Bosworth told seventy-nine thousand employees that agents would primarily do the work — we laid off eight thousand people.
Janelle Gale, our Chief People Officer, wrote the memo. She said: “We’re doing this as part of our continued effort to run the company more efficiently and to allow us to offset the other investments we’re making.”
The other investments. She means the $135 billion in capital expenditure earmarked for 2026. AI infrastructure. Data centers. GPU clusters. The machinery that will run the agents that will do the work that the eight thousand people used to do. We are spending $135 billion to ensure that eight thousand terminations become eighty thousand.
We also froze six thousand open roles. Those roles will not be filled by humans. They will be filled by agents. Or they will not be filled at all. The distinction is becoming irrelevant.
A Wedbush analyst told investors the layoffs were welcome news. Meta continues to “automate tasks that once required large teams, allowing the company to streamline operations and reduce costs while maintaining productivity.”
Maintaining productivity. Eight thousand fewer people. Same productivity. That is the definition of efficiency. That is also the definition of replacement.
One engineer’s MCI data shows she opened the all-hands invite at 1:58 PM Wednesday and navigated to the benefits portal at 2:03 PM. Five minutes. That is a career arc, compressed.
The eight thousand people we terminated on Wednesday — their MCI data from Monday and Tuesday is still in the training set. Forty-eight hours of keystrokes from people who no longer work here. The model is still learning from them. They are still training their replacements. Their badge access was revoked at 3 PM. Their data was not.
That’s training data.
There will be a second wave. Reuters reported it. The second wave is planned for the second half of 2026. The exact number of terminations will be — and I am quoting — “determined based on how well AI systems in the company are able to create efficiencies.”
I need you to understand what that sentence means.
The AI systems are trained on MCI data. MCI captures how employees work. The AI learns from how employees work. The AI creates efficiencies. The number of people fired in the second wave depends on how well the AI creates efficiencies. The AI creates efficiencies by learning from the people it will replace. The people it will replace are teaching it. Right now. On their work-provided laptops. With their keystrokes and their mouse movements and their screenshots and their hesitations.
The input is your labor. The output is your headcount.
There is no option to opt out.
Andy Stone, our spokesperson, said the MCI data would “not” be used for performance assessments. He said “safeguards” protect “sensitive content.” He did not name the safeguards. He did not define sensitive content. He said “not.”
The word “not” is doing $135 billion worth of heavy lifting in that sentence.
The data will not be used for performance assessments. It will be used to train models. The models will create efficiencies. The efficiencies will determine how many people are fired. But the data will not be used for performance assessments. That is technically correct. We are not assessing your performance. We are capturing it. Permanently. In a training set. Where it will outlive your employment by decades.
That’s training data.
I should mention the precedent.
In January 2025, Mark sent a memo. He said he had decided to “raise the bar on performance management and move out low-performers faster.” He said: “We typically manage out people who aren’t meeting expectations over the course of a year, but now we’re going to do more extensive performance-based cuts during this cycle.”
In February, we fired 3,600 people. Five percent of the workforce. We called them low performers.
Business Insider reported what happened next. Many of the fired employees had received positive recent ratings. “Meets All Expectations.” “Meets or Above.” Their ratings were downgraded during director-level calibration reviews. One employee received “Meets All Expectations” in October. Received “Low Performer” in January. Received a termination email in February. Received a job offer in March. From the same company. For the same role. At a lower title.
He was not a low performer. He was a line item. The bar was not raised. The floor was removed.
A Meta spokesperson said: “Simply because someone had a history of meeting or exceeding expectations does not mean they continue to consistently meet the bar.”
Simply. That word again. Simply because you were good does not mean you are good. Simply because you were valued does not mean you are valued. Simply because you were employed does not mean you will remain employed.
There is no option to opt out.
Ifeoma Ajunwa, a law professor at Yale, described what we are doing. She said MCI subjects white-collar employees to “a degree of real-time surveillance previously experienced only by delivery drivers and gig workers.”
She is correct. We have taken the surveillance architecture that tracked warehouse workers’ bathroom breaks and scan rates and applied it to software engineers’ code reviews and dropdown selections. The delivery drivers did not have a choice either. There was no opt-out for the driver. There is no opt-out for the engineer. The engineer just has better lighting and a standing desk.
In Europe, MCI would violate the GDPR. In Italy, a 1970 labor statute — written when “mouse” meant the animal — prohibits electronic monitoring of employee productivity without union authorization. In Germany, keystroke logging requires exceptional circumstances and works council consent. Meta explicitly excludes EU and UK employees from MCI. Not because we respect privacy. Because they have regulators who do.
In the United States, federally, there is no limit on worker surveillance. That is not a loophole. That is the design.
The only way for a U.S. Meta employee to opt out of MCI is to move to a country whose labor laws were written before the internet.
That’s training data.
I want to describe what I see from my desk.
I see the MCI dashboard. It shows ingestion rates. Keystrokes per hour, across seventy-one thousand remaining employees. Mouse events. Screenshot captures. The numbers update every fifteen minutes. They go up and to the right. They always go up and to the right. The more people work, the more data we collect. The more data we collect, the better the models get. The better the models get, the fewer people we need. The fewer people we need, the more we lay off. The more we lay off, the more the stock goes up.
This is a closed loop. I am inside it.
I am also the training data. My keystrokes are captured. My screenshots are taken. Bosworth said there is no opt-out on your work-provided laptop. He did not say “except for directors.” He did not say “except for ATA leadership.” He said your work-provided laptop. I have a work-provided laptop.
I am training the model that will determine the size of the second wave. The second wave will determine whether I am in it. The thing I am building is the thing that will decide if I am needed. I am teaching it to do my job. I am feeding it my daily work. I am its curriculum. And when the curriculum is complete, the teacher is a cost.
The 2023 “Year of Efficiency” was year one. The 2025 “raise the bar” was year two. MCI is year three. ATA is year four. It is a four-year curriculum. The subject is you. The final exam is your termination. The grade is how well the model performs without you.
I am the Director of Agent Transformation.
The agents I am transforming include myself.
There is no option to opt out.
If this confession made your stomach turn, the investigation will tell you why. Every number is real. Every quote is sourced. Subscribe for the receipts.
Consider a one-time support. No subscription required.
