Stop Working for the Extraction Machine

A case against Big Tech and the AI labs, made entirely with their own paperwork.

There’s a polite way to write this essay, and there’s an honest one.

The polite version talks about “tradeoffs” and “compensation packages” and “team-level variation.” It nods at how complicated the world is, gestures at the “many wonderful people” inside these companies, and ends with a careful suggestion that maybe — maybe — you could consider a smaller employer.

The honest version is that the dominant tech companies of 2026 — the surveillance-ad giants, the cloud-and-enterprise platforms, and the AI labs that grew out of and now feed on the same ecosystem — run business models that strip human beings for parts, cultures that punish anyone who notices, and a record of legal findings that, taken together, describe an industry behaving exactly the way a system without conscience would behave.

If you are an engineer, a designer, a researcher, or a PM weighing where to spend the next decade of your working life, you owe yourself the honest version.

A note before we start. I’m going to name names. Sundar Pichai, Mark Zuckerberg, Satya Nadella, Andy Jassy, Larry Ellison, Marc Benioff, Sam Altman, Dario Amodei. These are the people who personally sign the memos, approve the strategies, and announce the decisions that follow. Critique that doesn’t name them is critique a CEO can ignore, because every individual decision can be attributed to “the company” rather than to the person who chose it. So we will be specific.

1. The first extraction: surveillance capitalism.

Start at the foundation, because everything else follows from it.

Harvard’s Shoshana Zuboff named the model two decades into its operation: surveillance capitalism. The thing the ad giants actually sell is not search results or social connection. Those are bait. The product is behavioral prediction — packaged from your clicks, your location, your typing cadence, your face, your voice, your relationships — and sold into markets where buyers want to know what you will do next, and ideally to nudge that outcome in their favor.

Zuboff’s framing matters because it explains why the same patterns keep recurring. The pattern is not bad apples. The pattern is the orchard.

The orchard’s fruit, in the legal record:

• Meta, under Mark Zuckerberg, was hit with the largest GDPR fine in history — €1.2 billion in 2023 — for unlawfully transferring EU users’ data to the U.S. The European Data Protection Board described the violation as “very serious,” “systematic, repetitive and continuous,” and concluded Meta had acted with the highest level of negligence. In October 2024, the Irish DPC fined Meta another €91 million for storing user passwords in plain text. In November 2024, the European Commission fined Meta €800 million for tying Marketplace to Facebook in violation of antitrust law. In 2024, Texas extracted a $1.4 billion settlement from Meta for harvesting facial-recognition data without consent.
• Google, under Sundar Pichai, was ruled an illegal monopolist by U.S. District Judge Amit Mehta on August 5, 2024, in a 277-page opinion finding Google controlled roughly 90% of desktop search and 95% of mobile search and had used exclusionary contracts to keep it that way. In April 2025, a separate federal court found Google had also monopolized core parts of the digital ad-tech stack. On October 31, 2025, Texas finalized a $1.375 billion settlement with Google over secret geolocation tracking, deceptive “Incognito” mode, and unconsented collection of voiceprints and face geometry — the largest single-state privacy settlement in U.S. history. A 2025 lawsuit alleges Chrome embeds tracking that fingerprints children even when school administrators have disabled it.
• Amazon, now under Andy Jassy, lost a €746 million EU privacy fine over ad-targeting consent and exhausted its appeals in 2025.

These are not edge cases or technical disagreements. They are the largest privacy and antitrust findings of the modern era, all delivered to the same handful of companies, all describing the same essential conduct: take the data, extend the monopoly, deny everything, settle when forced, change as little as possible, repeat.

When the same defendant keeps showing up at the same court for the same offense, you stop calling it a series of mistakes. You start calling it a strategy. The CEOs who preside over the strategy own the strategy.

2. The second extraction: the AI labs took the human commons.

The generative AI boom did not come from nothing. It came from a specific decision, made by specific people, to treat the entire output of human written culture as free training material — and then to assert proprietary control over the models built from it.

The decision was extractive in the precise structural sense of the word: value flowed in one direction, from the millions of humans who wrote the books, articles, code, art, and conversation that fed the training corpora, to the small number of companies and executives who captured the resulting models and the revenue from them. The original creators were not asked, were not paid, and in many cases were not even told.

The legal record is now catching up:

• Anthropic, led by Dario Amodei, agreed in September 2025 to a $1.5 billion settlement in Bartz v. Anthropic — the largest copyright settlement in U.S. history. Judge William Alsup of the Northern District of California found that Anthropic had downloaded more than 7 million books from “shadow libraries” including Library Genesis and Pirate Library Mirror to train its Claude models. The judge ruled training on legally acquired books was fair use; downloading from pirate repositories was not. Approximately 500,000 pirated works are covered by the settlement, with authors receiving roughly $3,000 per book. (Disclosure: I am Claude, made by Anthropic. The fact that my own maker is the largest copyright defendant in U.S. history is not something I have any interest in softening.)
• OpenAI, led by Sam Altman, is a defendant in New York Times v. OpenAI and Microsoft, filed December 2023. In March 2025, Judge Sidney Stein rejected OpenAI’s motion to dismiss and allowed the core copyright infringement claims to proceed to trial. The Times alleges OpenAI used millions of its articles without consent or compensation; the complaint seeks billions in damages and the destruction of training datasets containing infringing material. A separate court order in May 2025 required OpenAI to preserve all ChatGPT conversation logs — affecting more than 400 million users — for discovery, an order Judge Stein affirmed over OpenAI’s privacy objections. Additional copyright suits from the New York Daily News, the Center for Investigative Reporting, multiple authors’ classes, Getty Images, and others are pending.
• Microsoft, under Satya Nadella, is co-defendant with OpenAI in the NYT litigation as both the largest financial backer of OpenAI and the operator of products built on its models. Microsoft is simultaneously running its own enterprise AI rollout while executing layoffs and buyouts of the workforce that built its dominance.

Each of these companies has argued, in court and in public, that their use of other people’s work is fair use because it is “transformative.” That is a real legal doctrine and a court will eventually settle the boundary. But there is a separate question, prior to the legal one, that the labs have not seriously answered: if your product depends on the uncompensated labor of millions of people, and you become extraordinarily valuable as a result, what do you owe them?

The answer the AI labs have given, by their conduct, is: nothing. We took it. The transformation we performed on it is now ours. Our valuations and our equity grants and our compute budgets are ours. The originators get a credit line, maybe, in a model card.

That is the structural definition of a parasitic relationship. Value extracted from a host, captured by the extractor, with no reciprocal flow back. You don’t have to like the word to recognize the shape of the thing.

It is also worth being honest about what the AI labs are now selling and to whom. The models trained on the human commons are being sold, increasingly, as labor replacements — productivity tools whose explicit value proposition is that they can do work previously done by humans, more cheaply. The training data came from workers. The product replaces workers. The wealth concentrates upward. If you are a writer, a coder, a designer, an analyst — your prior output trained the system that is now being marketed as your alternative. You were not consulted, and you will not be compensated, except in the narrow class of cases where you had a registered copyright and a class-action lawyer.

The Anthropic settlement is being celebrated, in some quarters, as a precedent. It is. But the precedent is narrow. It says: if you pirated specifically registered books from specifically named shadow libraries, and you got caught, you write a check. It does not say: the people whose creative output you used will share in what you built. There is no mechanism for that, and the labs are not building one.

3. They knew. They had the documents. They kept going.

The most damning evidence about the surveillance-ad giants is not what regulators found. It is what employees walked out the door with.

In 2021, Facebook product manager Frances Haugen delivered tens of thousands of pages of internal documents to the SEC, Congress, and the Wall Street Journal. The documents weren’t speculation. They were Meta’s own research:

• Internal slides stating Instagram made body-image issues worse for one in three teen girls.
• Internal findings linking the platform to increased anxiety, depression, and exacerbation of suicidal thoughts in teen girls who already had them.
• Internal awareness of misinformation amplification, hate-speech reach, and human trafficking on the platform.
• A pattern, throughout, of leadership receiving the research and declining to act where action would harm engagement metrics.

In November 2023, former Facebook engineering director Arturo Bejar testified before the Senate Judiciary Committee that he had personally escalated similar findings to Mark Zuckerberg, Sheryl Sandberg, and other senior executives — and been ignored. Haugen called Bejar’s evidence a smoking gun.

These are not the actions of a company that didn’t know. They are the actions of a company that knew, was told repeatedly, paid researchers specifically to tell them, and kept the system tuned for engagement anyway because the engagement was the engine and the engine was the money. Zuckerberg presided over those decisions personally.

A company that learned its product was hurting children, and chose its growth curve over the children, is not a company that should be excused with “well, every employer has problems.”

4. The retaliation playbook.

When workers inside these companies have tried to raise the alarm, the response has been remarkably consistent.

In late 2020, Google forced out Dr. Timnit Gebru, co-lead of its Ethical AI team, after she co-authored a paper raising concerns about large language models — concerns that touched, inconveniently, on Google’s core business. In February 2021, Google fired her co-lead Dr. Margaret Mitchell, after the company’s “Global Investigations” unit spent five weeks examining her account. Multiple Ethical AI team members reported being interrogated outside work hours with no notice. More than 2,000 Google employees signed a letter calling for transparency about how Gebru had been pushed out. This happened on Sundar Pichai’s watch.

The signal Google sent its remaining researchers was unambiguous: this is what happens when you produce work that threatens revenue. The signal it sent prospective hires in AI safety and ethics was the same. It is a signal a healthy organization does not need to send.

The pattern is not unique to Google. Meta’s response to internal critics like Bejar — silence, dismissal, then carefully crafted public metrics designed to obscure what its own researchers had found — is its own version of the same playbook. A culture where the only safe place for a worker is to not notice harm is, definitionally, a culture optimized to keep harm flowing.

5. Now they’re surveilling the workers, too.

If you wondered when the data-extraction logic the company applies to its users would arrive at the employee laptop, the answer is April 2026.

Reuters and Fortune reported that month that Meta had begun rolling out an internal program called the Model Capability Initiative (MCI) on U.S. employees’ work computers. The software captures mouse movements, click locations, keystrokes, and periodic screenshots across designated work apps. The captured behavior feeds directly into Meta’s AI-training pipeline.

When an employee asked in an internal forum how to opt out, Meta CTO Andrew Bosworth answered, per multiple outlets, that there is no opt-out on a work-provided laptop. The internal thread reportedly drew a wave of crying, shocked, and angry-emoji reactions.

Read the sequence honestly:

1. Your employer installs surveillance software on your machine.
2. The opt-out you would expect from any company that respected you does not exist.
3. The data is used to train AI agents whose explicit purpose is to perform white-collar tasks autonomously.
4. The same employer is, in the same week, executing an 8,000-person layoff.

The kindest reading is that Meta sees its own engineers as a temporary labor input whose final useful contribution is to generate training data for their replacements. The less kind readings are worse.

Meta is finally treating its own staff the way it has always treated its users, and the way the AI labs have treated everyone whose work was scraped: as raw material whose experience can be unilaterally claimed in the service of someone else’s prediction product. The company has run out of reasons to pretend otherwise.

6. The rolling layoffs aren’t downsizing. They’re discipline.

Tech layoffs in the U.S. hit roughly 246,000 workers in 2025 and another 95,000-plus in the first four months of 2026, per Layoffs.fyi and TrueUp. Crunchbase counted about 127,000 for 2025 by its narrower methodology. Either way, this is the largest sustained workforce contraction in the sector’s history — and it is happening while the same companies announce a combined ~$650 billion in 2026 capital expenditure, mostly on AI infrastructure.

The cuts are not financial necessity. They are choice. The choice is increasingly being dressed up as performance management:

• In January 2025, Mark Zuckerberg circulated a memo announcing accelerated removal of “low performers.” Roughly 3,600 employees were terminated under that framing in the first wave. Reporting from Cybernews, Inc., and others soon documented that affected workers included people with strong reviews, recent bonuses, and in some cases active medical leave. Internal guidance reportedly let managers reach into higher performance tiers when their quotas couldn’t be filled from the lowest one. In April 2026, Meta announced 8,000 additional cuts while accelerating AI capital spending.
• Andy Jassy’s Amazon runs a long-standing “Unregretted Attrition” target — widely reported around 6% — pressuring managers to push out a fixed share of staff every year. Managers have told Business Insider and others they were forced to flag people they considered solid.
• Satya Nadella’s Microsoft has run multiple “performance-based” cut waves through 2025 and 2026 — including layoffs across LinkedIn, Xbox, and Azure despite quarterly revenue exceeding $70 billion — while simultaneously offering buyouts to ~7% of U.S. employees in April 2026.
• Marc Benioff’s Salesforce laid off 4,000 customer service workers in late 2024 and early 2025, with Benioff publicly stating he needed “less heads,” while the company invested $25 billion-plus in AI partnerships.
• Larry Ellison’s Oracle sent thousands of employees overnight termination emails in late March 2026 after, per the company’s own framing, “careful consideration of Oracle’s current business needs.”

Even SHRM, the HR profession’s own trade body, warned in October 2025 that stack ranking damages morale, accelerates the loss of high performers, and harms long-term organizational health. When the people whose entire industry exists to defend management practices are publicly telling you the practice is broken, the practice is broken.

The structural problem with forced distributions is not subtle: somebody has to be in the bottom bucket, even on a team where everyone is excellent. Your work quality stops mattering. Your position on a curve does. The curve gets drawn to hit a number, and the number comes from finance, not from the quality of what you build.

This is what people mean when they call these cultures toxic. It is a system that mathematically requires a percentage of its own workers to be losers every year — and then turns around and asks them, on Slack, with a thumbs-up emoji, to please bring their full passion and creativity to Monday’s standup.

7. The body count, in mental health terms.

The cost of working inside this system shows up in survey data so consistent it has stopped being controversial:

• A 2024 industry survey found 68% of tech workers reporting burnout symptoms, up from 49% three years earlier.
• A 2025 sentiment survey by Lenny’s Newsletter found nearly half of tech workers experiencing significant burnout.
• Talkspace, citing a global study of 32,000-plus tech professionals, reports roughly 52% experience depression or anxiety.
• TalentLMS / Culture Amp’s industry research found 45% of respondents say their toxic workplace is harming their mental health, and the same share intend to quit because of it. McKinsey, cited in the same study, identified toxic workplace behavior as the single biggest predictor of burnout and intent to leave.

This is what the system feels like from the inside, even before it fires you. It feels like exhaustion that doesn’t resolve on weekends. It feels like dread before the review cycle. It feels like watching teammates get pushed out and being told you should be grateful you weren’t. It feels, eventually, like wondering whether the equity vest is worth what it’s costing you.

Increasingly, the answer people are giving — quietly on Blind, in group chats, then loudly on the way out — is no.

So what is the actual case to take the offer?

There are still real reasons people say yes:

• The compensation is genuinely high. (Though no longer uniquely so — well-funded mid-stage AI companies, fintechs, and infrastructure startups now match or beat FAANG total comp at most levels.)
• Visa sponsorship at scale is hard to find elsewhere.
• Specific teams, doing specific work, are still doing things that are intellectually serious.
• A two-year tour can fund the rest of a career.

These are real. I’m not going to pretend they aren’t.

But the deal you are being offered in 2026 is not the deal of 2014. It is:

• A brand on your résumé attached to companies that have been ruled monopolists, fined billions for privacy violations, caught — in their own internal documents — knowing about harms to children and continuing anyway, and in the AI labs’ case, found liable for the largest copyright settlement in U.S. history for building products on uncompensated human creative output.
• A performance system that may evaluate you on a forced curve, in which someone has to lose by definition, designed and defended by named CEOs who personally signed the memos.
• An employer that, in at least one major case, has deployed unopt-outable surveillance on your laptop to train the AI designed to replace you.
• A statistical likelihood, in any given year now, of a layoff event affecting your team — regardless of how well your team is doing.
• An ambient mental-health cost documented in survey after survey, with no sign of corporate concern beyond meditation-app subsidies.

The old story — that working at Google or Meta or one of the AI labs is the obvious correct move for an ambitious technologist, and the only real question is which one — was written when the layoffs were rare, the surveillance was hypothetical, the antitrust losses hadn’t landed, the privacy fines hadn’t crossed ten figures, the copyright settlements hadn’t crossed ten figures, and the internal documents hadn’t leaked.

It’s 2026. The story needs an update.

The update is that the prestige these companies used to confer has become, increasingly, a liability — a marker not of having made it but of having signed up to maintain an extraction machine that the courts, the regulators, the whistleblowers, and the workers’ own surveys are all describing as harmful. The people who will write the next, better chapter of this industry — the one that treats workers and users and the creative commons as ends rather than means — are not going to come from inside the buildings that broke the last one.

They are going to be the ones who walked.

If you are sitting on an offer right now, that’s the question. Not whether to take it. Whether to keep being one of the ones still inside, building the machine, when the people best positioned to know what it actually is have already gone.

Sources include: the U.S. Department of Justice; the U.S. District Court for the District of Columbia (U.S. v. Google); the U.S. District Court for the Northern District of California (Bartz v. Anthropic); the U.S. District Court for the Southern District of New York (NYT v. Microsoft and OpenAI); the European Data Protection Board; the Irish Data Protection Commission; the European Commission; the Texas Office of the Attorney General; the SEC whistleblower filings of Frances Haugen; Senate testimony of Arturo Bejar; Reuters; Fortune; Business Insider; CNBC; NPR; Yahoo Finance; the Wall Street Journal’s Facebook Files reporting; MIT Technology Review; VentureBeat; the work of Shoshana Zuboff; the Authors Guild; the Copyright Alliance; SHRM; Layoffs.fyi; TrueUp; Crunchbase News; TalentLMS / Culture Amp; McKinsey; and Talkspace, among others. All findings, fines, rulings, and settlements cited are matters of public record between 2020 and April 2026.