The Invisible Workforce

How Big Tech Breaks the Bodies and Minds of the Workers It Cannot See — and Then Disposes of Them In February 2026, a group of workers in Nairobi, Kenya, began speaking to Swedish journalists. Two newspapers — Svenska Dagbladet and Göteborgs-Posten — published the story that those workers, at enormous personal risk, had decided the world needed to hear. They worked for a company called Sama, a Nairobi-based outsourcing firm contracted by Meta. Their job was to review footage captured by users of Meta's Ray-Ban smart glasses — to annotate what was in the video clips, label objects and actions, and help train Meta's artificial intelligence systems. That is the polite, corporate description of what they were asked to do.

What they actually saw was something different.

During long ten-hour shifts, these workers were shown footage of people having sex, captured on camera without their knowledge. They saw people changing clothes in their bedrooms. They saw children filmed without consent. They saw individuals using toilets. They saw credit card numbers accidentally captured on camera as someone used their glasses to help a friend pay for something. This footage — intimate, private, often violating — came from users of Meta's glasses who had no idea that their daily lives were being sent to a labor force in Kenya to be catalogued, labeled, and used to train artificial intelligence. The workers who spoke out said they were not allowed to bring phones or any personal devices to work. Security cameras monitored their every move. The implication was clear: Meta knew this footage was sensitive enough to require containment of anyone who saw it. (Source: Gadget Review, "Meta Fires Contractors Who Exposed Ray-Ban Glasses Privacy Scandal," May 2026, https://www.gadgetreview.com/meta-fires-contractors-who-exposed-ray-ban-glasses-privacy-sca ndal; Android Headlines, "1,100 AI Trainers Were Fired After Blowing the Whistle on Meta's Ray-Ban Privacy Problem," May 2026, https://www.androidheadlines.com/2026/05/1100-ai-trainers-were-fired-after-blowing-the-whistle- on-metas-ray-ban-privacy-problem.html)

Less than two months after the story broke, Meta ended its contract with Sama. The Nairobi firm confirmed the termination of 1,108 workers — some receiving as little as six days' notice. Meta claimed Sama "did not meet its standards." Sama said it had consistently met all contractual expectations and had never been notified of any deficiency. The logic of what happened is not difficult to follow. Workers blew the whistle. Meta ended the contract. 1,108 people lost their jobs.

Meta has described the footage review as standard procedure covered by its terms of service. That is technically true, in the same way that every predatory agreement is technically disclosed in fine print that no one reads. The people whose intimate moments were shipped to Kenya to be watched by strangers during a ten-hour shift did not meaningfully consent to that. And the workers who had to watch it are now without income because they had the audacity to say so. (Source: OECD.AI Incident Report, "Meta's AI Smart Glasses Lead to Worker Harm and Privacy Violations in Kenya," April 30, 2026, https://oecd.ai/en/incidents/2026-04-30-7774)

This Is Not the First Time

The Kenya Sama story is new. The pattern it represents is not.

This is not the first time Meta has left a trail of traumatized and discarded workers in Nairobi. In January 2023, all 260 content moderators working at Sama's Nairobi hub — people who spent their working hours reviewing Facebook content for violence, child exploitation, and graphic material — were told their jobs were redundant. Forty-three of those workers filed a lawsuit for unlawful dismissal. A class action eventually grew to involve 185 workers. Dr. Ian Kanyanya, head of mental health services at Kenyatta National Hospital, assessed 144 of them. His findings, filed with Nairobi's employment and labor relations court in December 2024, were devastating: 81 percent of the workers assessed were suffering from severe PTSD. (Source: CNN Business, "Facebook inflicted 'lifelong trauma' on content moderators in Kenya, campaigners say, as more than 140 are diagnosed with PTSD," December 22, 2024, https://edition.cnn.com/ 2024/12/22/business/facebook-content-moderators-kenya-ptsd-intl)

One of the workers who led the protest — a South African man named Daniel Motaung — was fired for organizing. He sued Meta and Samasource Kenya, alleging forced labor, human trafficking, and union-busting. He told investigators that his team was forced to view hours of horrific content including beheadings and child sexual exploitation — for pay of less than $2.20 an hour. Over 80 labor organizations from across the world wrote to Meta demanding it stop attempting to silence him. Among the signatories was Meta's most high-profile whistleblower, Frances Haugen. (Source: Quartz Africa, "Meta faces pressure from rights groups to stop gagging Daniel Motaung," https://qz.com/africa/2191286/meta-faces-pressure-from-rights-groups-to-stop-gagging-daniel-m otaung)

This is a company valued at over a trillion dollars. It pays the people who protect the rest of its users from the worst content on its platforms less than the cost of a large coffee in San Francisco — and then terminates them when they complain about the trauma of doing it.

The Science of What This Work Does to Human Beings

I want to be precise about what content moderation does to a person, because the industry has perfected the art of keeping this invisible. These are not hypothetical harms. They are documented clinical realities.

Scientific peer-reviewed research confirms that content moderators — the people paid by technology companies to view graphic and distressing material so the rest of us don't have to — suffer at measurably higher rates from PTSD, anxiety, depression, intrusive thoughts, sleep disturbance, emotional detachment, and burnout. A cross-sectional survey published in 2024 found that 53 percent of self-identified content moderators reported clinical levels of general psychological distress. Research shows a dose-response relationship: the more traumatic content someone is exposed to, the worse their mental health outcomes. The exposure does not wear off when the shift ends. (Source: Research paper, "I've Seen Enough: Measuring the Toll of Content Moderation on Mental Health," arXiv:2511.09813, November 2024, https://arxiv.org/pdf/2511.09813; Cyberpsychology Journal, "The psychological impacts of content moderation on content moderators," 2023, https://cyberpsychology.eu/article/view/33166)

Researchers at SAGE Journals have characterized what happens to content moderators not as an accident of the job but as a structural feature of it: the commodification of traumatic vulnerability itself. In other words, the companies are not failing to protect these workers. They

are deliberately purchasing the capacity of these workers to absorb trauma — buying their psychological health as an input cost, the way you buy servers or storage or bandwidth. (Source: Amit Pinchevski, "Social media's canaries: content moderators between digital labor and mediated trauma," SAGE Journals, 2023, https://journals.sagepub.com/doi/abs/ 10.1177/01634437221122226)

The people doing this work in Kenya — and in Ethiopia, and in the Philippines, and in Indonesia, and in every other low-wage country where Big Tech has outsourced the burden of its most traumatic operational requirement — are not edge cases. They are a workforce of tens of thousands, doing labor that makes every social media platform you use every day possible. And they are doing it for wages, under conditions, and with mental health consequences that would be illegal to impose on workers in the countries where the profits are collected.

That is not an oversight. That is a business model.

The Amazon Warehouse: Blood, Sweat, and Next-Day Delivery

I want to move from the psychological labor of content moderation to the physical labor of fulfillment — because both tell the same story about how Big Tech treats the workers at the bottom of its supply chain: as a resource to be optimized until it breaks, then discarded.

In August 2015, the New York Times published an investigation — based on interviews with more than a hundred current and former Amazon employees — that described a workplace characterized by midnight emails, morning texts demanding to know why the midnight email had gone unanswered, employees crying at their desks, and workers experiencing health crises so severe they were brought out on stretchers. The reporters, Jodi Kantor and David Streitfeld, quoted a former employee saying: "Nearly every person I worked with, I saw cry at their desk." (Source: New York Times, "Inside Amazon: Wrestling Big Ideas in a Bruising Workplace," August 15, 2015, by Jodi Kantor and David Streitfeld, https://www.nytimes.com/2015/08/16/technology/inside-amazon-wrestling-big-ideas-in-a-bruisin g-workplace.html)

That investigation described the white-collar experience at Amazon's corporate offices. What has since been documented about the blue-collar experience in Amazon's warehouses is, if anything, worse — and far more physically immediate.

In December 2024, Senator Bernie Sanders released a 160-page Senate HELP Committee investigation into Amazon's warehouse safety record. Its findings were damning and detailed. Amazon's injury rate in 2022 was 6.6 percent — against an industry average of 3.2 percent. Amazon warehouse workers suffer serious injuries at a rate 2.6 times higher than their non-Amazon counterparts. An internal Amazon report — titled Project Soteria and cited by Senate investigators — found a direct relationship between the speed of tasks performed and the rate of worker injury. Amazon reviewed this evidence and chose not to act on it, according to Senate investigators, due to financial considerations. (Source: U.S. Senate HELP Committee, Amazon Investigation Report, December 2024, https://www.help.senate.gov/imo/media/doc/ amazon_investigation.pdf; CNN Business, "Senate report accuses Amazon of ignoring worker safety in productivity push," December 16, 2024, https://www.cnn.com/2024/12/16/business/ amazon-worker-safety-senate-report)

The most common injuries at Amazon warehouses are sprains, strains, and muscle tears — the kind of injuries that would sideline a professional athlete for weeks. The workers sustaining

them are not athletes recovering in specialized medical facilities. They are people who cannot afford to stop working, picking up shifts as Amazon's automated systems monitor their every movement, log every deviation from productivity targets, and generate disciplinary action for any worker who does not maintain the required pace. Injury rates spike during Prime Day, Black Friday, and Cyber Monday — exactly the moments when productivity pressure is at its highest and the human body is pushed furthest past its limits.

I remember, when I was at Amazon Web Services, the culture that the New York Times described was not a fiction. I saw it. I lived adjacent to it. I watched the machine demand everything and offer nothing in return except a paycheck and the expectation of gratitude. The corporate employees at least had the dignity of a salary and a desk and the option to update their résumé. The warehouse workers have a scannable wristband and a productivity quota and a body that is slowly being consumed by the pace of next-day delivery.

The Death of Main Street by Algorithm

There is a third category of harm that Amazon has inflicted — quieter than the warehouse injuries, less dramatic than the crying at desks, but perhaps more structurally devastating to the communities that make up the actual fabric of American life. It is what has happened to the small businesses that once lined the main streets of every town in this country.

In 2016, according to the e-commerce research firm Marketplace Pulse, Amazon took roughly one-third of every dollar made by merchants selling on its platform, in the form of fees and advertising. By 2022, that number had surpassed 50 percent. By 2024, Amazon had collected over $150 billion in revenue from third-party seller fees alone — a number so large it would rank inside the Fortune 25 if it were its own company. (Source: Fortune, "Amazon's record earnings contain a mystery: How much are fees from small sellers to thank?" February 7, 2025, https://fortune.com/2025/02/07/amazon-2024-earnings-third-party-seller-fees-profits/)

When you combine referral fees, fulfillment costs, advertising spending, and storage charges, the total bite Amazon takes from a seller's revenue runs between 30 and 50 percent — often more. For small sellers with already thin margins, that arithmetic is not a business model. It is a slow execution. Independent analysis confirms that for many product categories, the cost of selling on Amazon now exceeds any conceivable profit margin for a small producer. The platform that was supposed to give every small business a global storefront has become a mechanism for extracting value from the producers of goods at a rate that would have been recognized immediately in any other era as monopolistic extraction. (Source: SmartScout, "The Top Challenges Facing Amazon Third-Party Sellers in 2025," April 2025, https:// http://www.smartscout.com/blog/amazon-third-party-sellers-problems)

I grew up in the Bay Area. I remember what a main street looked like before Amazon. I remember the hardware store where the owner knew your name. The bookshop where someone could recommend something you had never heard of. The clothing store run by a family that had been in the neighborhood for twenty years. I remember summer jobs — the kind of entry-level, community-embedded work that gave a teenager their first paycheck, their first experience of responsibility, their first relationship with a neighborhood business that was also a neighborhood institution.

Those jobs are gone. Those stores are closed. The people who owned them could not compete with a company that uses algorithmic pricing, logistics infrastructure built with government subsidies, and a marketplace model that takes half of every dollar — and then uses the data from that marketplace to identify which products to manufacture under Amazon's own private label, in direct competition with the sellers whose fees funded the platform's growth. This is not free-market competition. It is a rigged table at which only one player can never lose.

And the workers who replaced the shop owners — the people packing and shipping the goods in the warehouses — are not getting what the shop owners had. They have no ownership. No relationship with the customer. No stake in the product. They have a productivity quota, a scannable badge, and a body that is being systematically injured at 2.6 times the industry rate while Amazon posts record profits and its founders build rockets.

What Comes Next: Robots, and the Illusion of Concern

Amazon has announced with considerable fanfare its investments in warehouse robotics and automation — framing them as safety improvements for workers. There is a meaningful sense in which reducing the number of times a human being lifts a fifty-pound box by replacing the lift with a machine is a genuine improvement. I will acknowledge that straightforwardly.

But let me also say what is not being acknowledged: the same automation that reduces some physical strain will eliminate most of the jobs. Amazon has been working toward fully automated fulfillment for years. When those jobs go — and they will go — the communities that built their economic identity around the Amazon warehouse on the edge of town will not receive a severance package from the algorithm that replaced their neighbors. They will receive nothing. The workers who spent years being injured at twice the industry rate building Amazon's logistics empire will be thanked for their service and shown the door.

This is the endpoint of a business model that has always treated workers as an input cost to be minimized and replaced. First, it minimized wages by offshoring content moderation to Kenya. Then it minimized labor costs by squeezing warehouse workers at speeds that break their bodies while monitoring every second with surveillance technology. Now it is minimizing the cost of labor entirely, by replacing human beings with machines — not because the workers were inadequate, but because the machines are cheaper.

The question that no quarterly earnings call will ever ask is: what do we owe the people whose labor we consumed on the way to not needing them anymore? What do we owe the communities whose main streets we closed, whose summer jobs we eliminated, whose local economies we hollowed out — on the way to building the most profitable logistics company in human history?

The White-Collar Harvest: When the Employee Becomes the Dataset

Amazon took the bodies of its warehouse workers and broke them. The injury rates, the productivity quotas, the stretchers, the Senate investigation — all of it is documented above. The endpoint of Amazon's model is no longer speculative. It is a fully automated warehouse. The workers who built it will be thanked for their service and shown the door. The communities around those warehouses will receive nothing.

What Meta announced on April 21, 2026 is the same story — moved upstairs. Into the glass offices. Onto the corporate laptops. Into the hands and minds of the engineers, product managers, and designers who believed, because they had degrees and stock options and

ergonomic chairs, that they were a different category of worker than the people packing boxes in New Jersey.

They are not. They never were. The extraction just took longer to reach them.

On April 21, 2026, Meta disclosed to its U.S.-based employees — in an internal memo posted to a channel belonging to the company's Meta Superintelligence Labs team, obtained by Reuters and CNBC — that it had installed tracking software on their corporate computers. The program is called the Model Capability Initiative, or MCI. It captures mouse movements, clicks, keystrokes, and periodic screenshots of employees' screens. It runs continuously across a designated list of work-related applications and third-party platforms. The data feeds directly into Meta's AI training pipeline — used to build AI agents capable of performing, without human involvement, the same white-collar tasks the employees are being recorded performing. (Source: Reuters, "Meta tells employees it's tracking their keystrokes to train AI," April 21, 2026, https://www.reuters.com/technology/meta-tells-employees-its-tracking-their-keystrokes-train-ai-2 026-04-21/; CNBC, "Meta is tracking employee keystrokes on Google, LinkedIn, Wikipedia as part of AI training initiative," April 22, 2026, https://www.cnbc.com/2026/04/22/meta-tracks-employee-usage-on-google-linkedin-ai-training-p roject.html)

The list of platforms being monitored was widely circulated inside Meta. It includes Google, LinkedIn, Salesforce's Slack, Microsoft's GitHub, Atlassian, and Meta's own properties. Notably, the list originally included OpenAI's ChatGPT and Anthropic's Claude — the two most widely used AI assistants in professional settings — before those platforms were quietly removed from the tracking list without explanation. The inference is not difficult to draw: capturing how Meta's own employees interact with competitor AI systems would expose the company to legal, competitive, and reputational consequences it was not prepared to absorb. The employees themselves received no equivalent protection. (Source: CNBC, ibid.)

What MCI Can Actually See — and Why That Should Terrify You

I want to stop here and ask a question that the corporate communications around MCI have been careful not to ask directly. What does a program that captures mouse movements, keystrokes, and continuous screenshots of an employee's screen throughout their working day actually see?

Not in the abstract. In practice.

It sees every email a worker receives from their doctor's office — appointment reminders, test results, treatment updates, prescription confirmations. It sees the pharmacy order a worker places online during a lunch break for a medication they would prefer their employer know nothing about. It captures the name of that medication, the dosage, the refill date, and the condition it treats — whether that is depression, HIV, diabetes, a psychiatric disorder, a chronic illness, or a pregnancy. None of that information was offered to Meta voluntarily. None of it is relevant to AI model training. All of it is now in a training pipeline.

It sees the text message a worker sends from their computer to their spouse about a child's diagnosis. It sees the insurance portal a worker visits to check whether a procedure is covered. It sees the search queries a worker runs to understand what a new diagnosis means, or what a medication's side effects are, or whether their condition qualifies for disability accommodation. It sees the moment a worker reads a piece of news that terrifies them — about a policy change, a

drug approval, a clinical trial — and the private, immediate, human response they type to someone they trust.

It sees the bank notification that appears on screen when a worker's direct deposit arrives. It sees the credit card statement a worker opens to dispute a charge. It sees the financial account a worker accesses to check whether they can afford the treatment their doctor just recommended. It captures account numbers, transaction histories, and the particular combination of financial stress and medical need that defines what is happening in a person's life at any given moment.

It sees the political article a worker reads during their lunch break. It sees whether that article is from a conservative outlet or a progressive one. It sees the social media post a worker opens and reads in full — including the political content, the party affiliation of the person who wrote it, and whether the worker lingers, shares it, or types a response. In an era when employers screen candidates for political beliefs, when immigration authorities monitor social media, and when political affiliation can affect professional relationships in ways that are difficult to document but impossible to ignore — this is not a trivial capture. It is a profile.

It sees the personal email from a family member describing a health crisis — a parent's cancer diagnosis, a sibling's addiction, a child's mental health hospitalization — landing in a worker's inbox at two in the afternoon while they are trying to finish a product brief. It sees the response the worker types. It sees the web searches they run afterward. It sees the grief happening in real time, on a corporate device, because human beings do not cleanly separate their lives from their work hours no matter how many employee handbooks tell them to try.

None of this was supposed to be a data collection exercise. All of it now is.

Meta's Response: "Don't Check Personal Email on Your Work Computer"

When Meta employees raised exactly these concerns in internal discussions — how would the company prevent the capture of health information, financial data, immigration status, family communications — Meta CTO Andrew Bosworth's response was captured in internal communications obtained by Platformer. His answer, in full: "Gmail is an approved context so if you have concerns it may be best not to check personal email on your work computer." (Source: Platformer, "The week that Meta employees became training data," April 2026, https://www.platformer.news/meta-mci-monitoring-layoffs-knowledge-work/)

Read that again.

The people whose health records, financial information, family communications, and political views are now being captured to train Meta's AI were told, by their company's Chief Technology Officer, that their only recourse was to stop using their work computers for anything personal.

This is not a privacy protection. It is a liability shift. It is Meta telling its employees: we have installed surveillance infrastructure on your machine that will capture everything visible on your screen, and the responsibility for ensuring your most private information does not enter our training pipeline belongs to you. Not to us. You.

That answer is not just callous. It is legally and ethically incoherent. Workers in the United States have well-established rights under federal and state law that govern what employers can and cannot collect about their medical conditions and health status. The Americans with Disabilities Act, the Health Insurance Portability and Accountability Act, the Genetic

Information Nondiscrimination Act — these are not aspirational documents. They are federal law. They exist precisely because Congress recognized that the power imbalance between an employer and an employee, left unchecked, will always resolve in the employer's favor — and that certain categories of personal information are too sensitive, too intimate, and too potentially weaponizable to be left to the goodwill of the party with the greater power.

Meta has now built a system that captures exactly this category of information — not deliberately, perhaps, but systematically, inevitably, and continuously — and has told its employees that the solution is behavioral self-regulation on the part of the surveilled. The company that cannot be trusted to build a facial recognition system ethically, that terminated 1,108 Kenyan workers for speaking about what they saw, that complied with 94 percent of Israeli government censorship requests — this company now holds a continuously updated, screenshot-by-screenshot record of its employees' medical lives, financial lives, political lives, and family lives, and assures them it will not be misused.

I am not reassured. Neither, judging by internal discussions described as "dystopian" by multiple employees, are they. (Source: CNBC, ibid.)

The Political Dimension: What Your News Feed Says About You

Let me be specific about the political surveillance dimension of MCI, because it has received less attention than the medical and financial angles — and it is potentially the most dangerous of all.

MCI runs on Google. It runs on LinkedIn. It runs across the web browsing that happens on an employee's corporate device throughout the day. In practice, that means it captures what news sources a worker reads, which political stories they open and read to completion, which they skip, which social media posts they engage with, and what they type in response to political content.

This is a comprehensive political profile — assembled not from declared opinions or party registrations, but from the actual behavioral fingerprint of how a person consumes political information in private moments throughout their working day. It captures the difference between a worker who reads only mainstream centrist news and one who reads independent progressive journalism. It captures who reads articles about union organizing, about immigration enforcement, about the political affiliations of Silicon Valley executives. It captures, in other words, exactly the information that employees in any politically sensitive environment have the strongest interest in keeping private from their employer.

The memo's assurance that this data "will not be used for performance assessments" deserves exactly the skepticism it has received. Meta has a documented history of making assurances about data use and then violating them. Its terms of service once said it would not sell user data. The Facebook Papers documented that internal decisions routinely prioritized engagement and profit over the safety policies Meta publicly committed to. The company that fired 1,108 Kenyan workers for speaking to journalists about what they saw on their work screens is asking its American employees to trust that their keystroke data will remain in the training pipeline and go nowhere else.

History does not support that trust. The data is the value. The data will be used.

The Superintelligence Memo: The Organizational Architecture Behind MCI

MCI did not appear in a vacuum. It is the operational expression of a corporate restructuring that Zuckerberg announced in the weeks before — and which reveals precisely how Meta is treating the data supply problem at the core of its AI ambitions.

In a memo released in full by CNBC, Zuckerberg announced the formation of Meta Superintelligence Labs, or MSL — a new unit consolidating all of Meta's AI model development, product teams, and fundamental AI research under single command. To lead it, Zuckerberg named Alexandr Wang — the 28-year-old former CEO of Scale AI, which Meta acquired a 49 percent stake in for $14.3 billion — calling him "the most impressive founder of his generation" and creating for him the new role of Chief AI Officer. Wang's appointment was not ceremonial. Scale AI built its entire business on one specific capability: harvesting workflow data from contractors to train AI systems. Wang said in 2024: "For a lot of the capabilities that we want to build into the models, the biggest blocker is actually a lack of data. There's no pool of really valuable agent data that's just sitting around anywhere. And so we have to figure out how to produce really high quality data." (Source: Entrepreneur, "Mark Zuckerberg Reveals Meta Superintelligence Labs," https://us.entrepreneur.com/business-news/mark-zuckerberg-reveals-meta-superintelligence-lab s/ 494047; Platformer, ibid.)

MCI is how Meta produces that data. And the workers generating it are, in structural terms, Meta's version of Scale's contractor workforce — except that these workers have salaries, stock options, and the expectation that their employer would not record their screen in perpetuity and feed the result into a model designed to eliminate their jobs.

Zuckerberg committed up to $135 billion in capital expenditure for 2026 to build Meta's AI infrastructure. At the same time, the company confirmed it would lay off ten percent of its workforce — approximately 8,000 people — while declining to fill an additional 6,000 open positions. These announcements came in the same week as MCI. (Source: Fortune, "Meta will start tracking employees' screens and keystrokes to train AI tools," April 21, 2026, https://fortune.com/2026/04/21/meta-will-start-tracking-employees-screens-and-keystrokes-to-tr ain-ai/)

The data collection and the displacement are not parallel stories. They are the same story, in sequence. Capture the knowledge. Train the model. Reduce the headcount. Repeat.

Zuckerberg offered his vision on the earnings call that preceded these announcements. "We're starting to see projects that used to require big teams now be accomplished by a single very talented person," he told investors. That sentence is the operating manual for everything that follows. The talented person's cognitive process is being recorded. The AI trained on it will eventually perform that process without the person. The compliment is the announcement of the mechanism. And the employees being complimented are funding it with every keystroke.

On a separate internal memo, Meta CTO Andrew Bosworth described the destination with equal clarity: "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." He called it a "closed loop" — agents that would "automatically see where we felt the need to intervene so they can be better next time." (Source: Detroit News, "Meta employee mouse movements, keystrokes, AI training data," April 21, 2026, https://eu.detroitnews.com/story/tech/2026/04/21/metaemployee-mouse-movements-keystrokes -ai-training-data/89717625007/)

The people whose keystrokes are being recorded to train the agents are, in Bosworth's own words, the people whose jobs the agents are designed to perform. Let no one pretend otherwise.

The Pattern Completed: From Warehouse to Workstation, from Body to Mind

I want to be precise about what distinguishes MCI from every earlier form of worker surveillance documented in this chapter.

Amazon's monitoring of its warehouse workers is about controlling the pace of physical labor. The scannable wristbands, the productivity quotas, the algorithmic discipline systems — they are designed to maximize output per body per hour, until the body is replaced by a machine. What is extracted is physical energy. What is broken is the body — the tendons, the backs, the rotator cuffs of people packing boxes at speeds that Amazon's own internal report documented as directly causing injury.

Meta's MCI extracts something different and something more intimate: the cognitive process of a human mind at work. The click sequences that navigate a complex engineering decision. The keyboard shortcuts that represent years of accumulated professional knowledge. The decision trees embedded in how an experienced product manager moves through the tools of their daily work. The correspondence with a doctor that happens to land on screen during that work. The credit card statement opened in a browser tab. The political article read during a lunch break. The text from a family member about a health crisis.

This is not energy being measured. This is a human life being digitized — the professional parts and the personal parts equally, because human beings conducting their working day on a machine do not sort themselves cleanly into categories that a tracking program can separate. Their private life follows them to work. The machine captures everything it can see.

The Amazon warehouse worker whose body is consumed in the service of next-day delivery receives no share of the logistics empire their labor built. The Meta engineer whose cognitive process — and medical history, financial data, and political views — is captured and fed into an AI training pipeline receives no equity in the agent their keystrokes trained. In both cases, the worker is the raw material. In both cases, the worker is told to be grateful for the opportunity to contribute. And in both cases, the endpoint is the same: the worker is made redundant by the very product their labor produced, while their most private information remains in a pipeline they can no longer access or control.

The content moderators in Nairobi are the invisible workforce of Meta's past — low-wage, offshore, discarded when they became inconvenient. The engineers and product managers subjected to MCI are the invisible workforce of Meta's present — well-compensated, American, and equally powerless once the extraction is complete. The geography is different. The salary is different. The chairs have better lumbar support. The structural position is identical: they are an input cost to be minimized and a data resource to be harvested, and those two facts exist in sequence, not in conflict.

Let me be direct. My medical condition is none of Meta's business. What medication I take during my working hours is none of Meta's business. The credit card number that appears on my screen when I pay a bill during lunch is none of Meta's business. The political article I read, the news source I prefer, the party affiliation I may or may not hold — none of it is the property of

the company that installed software on my machine and told me the solution was to stop checking my personal email at work.

This is not a terms of service dispute. It is a violation of the basic human right to a private life — a right that does not pause when you sit down at your desk in the morning and resume when you log off at night. The law in this country has historically recognized that. The question is whether the law will move fast enough to catch what Meta is doing before the data is already trained into a model that cannot be untrained.

A fool with a tool is still a fool. A company that harvests its employees' minds, their medical records, their financial lives, and their political identities to build the machines that replace them is something more deliberate than foolish.

It is operating exactly as designed.

Sources:

Reuters, "Meta tells employees it's tracking their keystrokes to train AI," April 21, 2026: https:// http://www.reuters.com/technology/meta-tells-employees-its-tracking-their-keystrokes-train-ai-2 026-04-21/

CNBC, "Meta is tracking employee keystrokes on Google, LinkedIn, Wikipedia as part of AI training initiative," April 22, 2026: https://www.cnbc.com/2026/04/22/meta-tracks-employee-usage-on-google-linkedin-ai-training-p roject.html

Fortune, "Meta will start tracking employees' screens and keystrokes to train AI tools," April 21, 2026: https://fortune.com/2026/04/21/meta-will-start-tracking-employees-screens-and-keystrokes-to-tr ain-ai/

Platformer, "The week that Meta employees became training data," April 2026: https:// http://www.platformer.news/meta-mci-monitoring-layoffs-knowledge-work/

Detroit News, "Meta employee mouse movements, keystrokes, AI training data," April 21, 2026: https://eu.detroitnews.com/story/tech/2026/04/21/metaemployee-mouse-movements-keystrokes -ai-training-data/89717625007/

Entrepreneur, "Mark Zuckerberg Reveals Meta Superintelligence Labs": https://us.entrepreneur.com/business-news/mark-zuckerberg-reveals-meta-superintelligence-lab s/494047

ArtVoice, "Meta Is Recording Its Employees' Keystrokes And Mouse Clicks To Train AI That Could Eventually Replace Them," April 22, 2026: https://artvoice.com/2026/04/22/meta-is-recording-its-employees-keystrokes-and-mouse-clicks-t o-train-ai-that-could-eventually-replace-them/

The Social Obligation That Has Been Abandoned

Every major technology company that has built its empire on human labor has made, implicitly or explicitly, a bargain with the society that makes its success possible. That bargain goes something like this: you give us your labor, your data, your consumer spending, your infrastructure, and your regulatory tolerance. In return, we give you jobs, economic participation, and a share — however modest — in the prosperity we generate.

Amazon has broken that bargain. Meta has broken it. Every Big Tech company that outsources its most traumatic labor to low-wage countries, pays below living wages, monitors workers at granular surveillance levels, maintains injury rates that would be considered unconscionable in any other industry, and then prepares to automate the jobs away while posting record profits — has broken it.

The shareholder is not the only stakeholder. This is not a radical idea. It is the foundation of every labor law, every workplace safety regulation, every minimum wage, every right to organize that workers have ever fought for and won. A company exists in a community. It uses that community's roads, its water, its educated workforce, its legal system, its courts — all of the public infrastructure that makes commerce possible. It owes that community something in return.

What Amazon owes the workers in its warehouses is not just a paycheck. It is working conditions that do not break their bodies. It is injury rates that are not 2.6 times the industry average. It is an honest accounting of what automation will do to their livelihoods, and a plan for transition that treats them as human beings rather than units of labor whose productive life has expired.

What Meta owes the content moderators it has hired through contractors in Nairobi is not just a termination with six days' notice. It is mental health care proportionate to the trauma it purchased. It is compensation for the psychological harm it imposed deliberately, systematically, and profitably. It is a basic acknowledgment that the people it hired to watch beheadings and child sexual exploitation and intimate footage captured through smart glasses are not disposable components of a data pipeline but human beings whose suffering was a cost that Meta chose to offload rather than internalize.

The workers cannot be the ones who absorb the externalities of Big Tech's profits. That is the argument this chapter is making. It is not a new argument. It is the oldest argument in the history of labor. But it has to be made again — because the companies making the decisions have convinced themselves, and many of their shareholders, that it no longer applies to them.

It applies to them. It will always apply. The arc of accountability, like the arc of history, is long. But I have seen enough to know it bends.

--- The workers I have just documented are made invisible by design. So are other things. In Chapter Nine, I turn to a different kind of erasure — not of labor and bodies, but of voice and evidence. The same platforms that extract value from workers in Kenya and warehouses in New Jersey have simultaneously erased the testimony of an entire people under siege. What follows is the documented record of how Meta, TikTok, X, YouTube, and LinkedIn silenced Palestine — and why that silencing is not a bug in the system, but one of its most consistent features.