Meta spent this spring hunting for an edge in the AI race and found one wearing a company lanyard: its own employees. The company wired the laptops of tens of thousands of U.S. workers to record their keystrokes, mouse movements, and screen content — a program it called the Model Capability Initiative — so its models could learn to do white-collar work by watching the very people whose jobs it was about to automate. The workers were told there was no opting out. They named the little icon that flashed on whenever the recorder switched on “the Eye of Sauron,” and then they did the one thing nobody in Menlo Park saw coming: they organized, they signed, and they killed the program.
The full story is now laid bare in Business Insider’s reporting, part of a new investigation into the AI data grab — the race by every major lab to mine anything a human being has ever touched. It is a story about what happens when a company decides its workers are raw material, and learns too late that raw material can print flyers and circulate a petition.
Andrew Bosworth — The CTO Said ‘No Opt-Out,’ Then Watched His Own Workers Beat Him
Andrew Bosworth, Meta’s chief technology officer, became the face of the program and the lightning rod for the fury it drew. When it launched, the message to staff was blunt: there would be no opting out. The announcement drew an immediate wall of angry reactions, and the top comment beneath it was four words — “this makes me super uncomfortable.” When a cybersecurity engineer pressed Bosworth on how the data would be protected and accused the AI division of treating security as an afterthought, Bosworth swatted the concern away and insisted the data was tightly controlled.
The irony would not land for weeks. In a leaked recording later published by the activist group More Perfect Union, Bosworth said — three separate times — that the workers who fought the program had made the entire company worse, not better. He was certain the petition signers thought they were helping. They weren’t, he insisted; they had only made it harder to change things from the inside.
A CTO who can’t hear his own workforce is not a leader. He’s a landlord explaining why the tenants shouldn’t complain about the cameras.
The Workers — They Named the Machine Watching Them, Then Signed Their Names Against It
The resistance started the way these things always do: with jokes, then memes, then a name. Employees circulated anti-surveillance memes comparing management to the omniscient director in The Truman Show, and they called the recording icon “the Eye of Sauron.” The joke was the first crack in the fear, and it did what humor always does — it made the thing small enough to fight.
Then came the petition. More than 1,800 workers signed a demand that Meta stop harvesting their computer-activity data — software engineers, research scientists, designers. The signers spelled out their fears in plain terms: that the recorder would swallow passwords, immigration paperwork, credit card numbers, the private life of a laptop. One group of activists went further, spreading flyers across Meta’s campuses that carried the language of the National Labor Relations Act, reminding their colleagues that speaking out against workplace conditions is legally protected.
Meta’s first concession was a 30-minute pause button — a crumb. The activists called it insufficient and demanded the program be scrapped entirely, immediately, forever. They did not blink.
The Leak — ‘Not So Tightly Controlled’ Was the Warning Nobody Believed Until It Happened
Then the warning came true. In late June, an engineer discovered that the data the program had collected — full prompts, transcriptions, private conversations, people and performance data — had been viewable across the entire company all along. The discovery went viral internally under a headline that quoted Bosworth’s own earlier reassurance back at him, dripping with sarcasm. “Hey Metamate,” one employee wrote, “define the word ‘hubris’ for me.”
Within hours, Meta switched the program off. It insisted the program was suspended, not canceled, and that any future version would be optional. Two months later, the company says it is still suspended, and it declined to comment further for the investigation.
The containment worked right up until the data was sitting where it shouldn’t have been.
Mark Zuckerberg — He Wanted More Data; His Own People Handed Him a Lesson Instead
The man at the top had been clear about why the program existed. Mark Zuckerberg told employees that AI models learn by “watching really smart people do things.” The Model Capability Initiative was that philosophy made literal: a machine watching the smartest people in the building, click by click, so it could eventually do their jobs without them.
The workers saw the math before the executives did. Some said the screen capture made the risk of being replaced by AI feel real for the first time — a “big reality check,” as one organizer put it. And so the people Zuckerberg was trying to turn into training data did the most human thing imaginable: they refused to be quietly mined.
The episode lands at the front of a wider fight. Across the industry, workers are starting to realize their own activity is the most valuable untapped dataset left, and they are pushing back. The same season Meta was forced to retreat, former Spirit Airlines flight attendants briefly blocked Google from buying the bankrupt airline’s data. The pattern is the same: the companies want the data, and the people who generate it are learning to say no.
Whether the data was ever even useful is its own punchline. One manager said it never made it into Meta’s latest models. Bosworth later said it was more valuable than expected. Either way, the program is dead, and the workers who killed it have been told by their own CTO that they made the company worse.
They didn’t make the company worse. They made it answerable.
Sources: Business Insider, The Guardian, Reuters, CNET.