How Meta tried, and failed, to grab its employees' data for AI

This spring, members of Meta’s elite AI unit, TBD, decided that the social media giant needed an edge in the AI race.

Data from tens of thousands of Meta employees.

Meta introduced a program to collect data from employees for AI training, called Model Capability Initiative (MCI).

Employees’ keystrokes and mouse movements could help Meta’s AI replicate how humans use computers, a critical task as AI labs race to automate white-collar work like coding and accounting.

But Meta had a very human problem: its own employees, who revolted over the mandatory project as soon as it was announced.

Their resistance took the form of flyers, memes, and a viral petition against what some characterized as “the Eye of Sauron.” Organizers say that their constant pressure forced Meta to make concessions and eventually drop the program.

Corporate data has become a gold mine for training AI because it helps algorithms deal with messy, real-world work environments — and workers are increasingly fighting back.

Former Spirit Airlines flight attendants temporarily blocked Google from buying the bankrupt airline’s database for $10 million.

Over 1,800 Meta employees signed a petition calling for the company’s internal tracking to be stopped.

Meta halted the tracking program after an internal data leak this summer, saying it would switch to being opt-in only.

Two months later, Meta says the program remains suspended.

The company declined to comment further for this article.

This story is part of a new series called The AI Data Grab, examining how companies are racing to collect as much data as possible to build the world’s most advanced AI systems.

For this story, Business Insider interviewed half a dozen current and former Meta employees, and reviewed internal posts and other materials to uncover new details about how the program began, how workers fought it, and the leak that brought it to a halt.

The MCI program originated within Meta’s AI unit, called TBD, a group of about 100 researchers widely considered the tip of the spear of the tech giant’s AI efforts.

One of the leaders of the effort was Rylan Schaeffer, three people told Business Insider.

Schaeffer reported to two people at the apex of Meta’s AI efforts: its AI chief, Alexandr Wang, and a top researcher, Jack Rae.

According to his LinkedIn, Schaeffer left Meta in July to work on something new.

He didn’t respond to requests for comment from Business Insider.

TBD members rarely interact with Meta’s legions of other employees, working in a separate area within Meta’s Menlo Park campus, according to two people familiar with the matter.

Their online workplace profiles are sparse, displaying few details beyond their name — unlike more typical profiles, which showcase people’s managers and what they’re working on.

One April morning, Meta announced the MCI program to the wider workforce. “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,” an internal announcement read.

The announcement drew an immediate backlash and a slew of angry-face emojis after Meta’s chief technology officer, Andrew Bosworth, said there would be no opting out. “This makes me super uncomfortable” was the top-rated comment in response to the announcement.

Bosworth would become a lightning rod inside Meta for his staunch defense of MCI.

He touted its strict privacy guardrails after a cybersecurity engineer asked how Meta would secure the data the program collects and accused the AI division of “moving fast and implementing security as an afterthought.” Bosworth replied, “That isn’t remotely a fair characterization!

How would any of this end up in ‘the open’?

This is tightly controlled data.” The intense online backlash spurred a grassroots movement from Meta employees.

Workers feared MCI would capture sensitive data that US workers plugged into their laptops, such as passwords, immigration forms, and credit card numbers, two organizers told Business Insider.

Some workers said they also feared that the screen capture data would be used to replace scores of Meta employees with AI — a risk many hadn’t taken that seriously before, a third organizer said, calling it a “big reality check.” Employees posted anti-AI memes in an internal Meta employee channel.

One meme seen by Business Insider portrayed Meta’s management as the omniscient villain, Christof, from “The Truman Show.” Employees jokingly dubbed the icon that appeared whenever MCI recorded their screens “the Eye of Sauron.” An anti-MCI petition gained steam, garnering over 1,800 signatures.

Staffers spelled out the link to the MCI petition on a building at Meta’s Burlingame office.

A small group of activists within Meta spread flyers around campuses, making sure they included language from the National Labor Relations Act that speaking out against workplace issues is protected, one of the organizers said.

In what the activists viewed as an early concession, Meta allowed employees to temporarily opt out of the tracking program for up to 30 minutes in June, during which they could perform more sensitive tasks they didn’t want ingested by an AI model.

It wasn’t enough for Meta’s anti-MCI activists, who called it “insufficient and unsatisfactory” in a statement to Business Insider at the time. “We demand, and will continue to demand, that Mark and his leadership team scrap MCI entirely, immediately, forever,” they wrote, referring to Meta CEO Mark Zuckerberg.

Soon, Meta’s own security decisions would force its leadership to take drastic action.

In late June, Meta engineers woke up to an internal post that had gone viral among employees.

An engineer had discovered that data collected by MCI had been viewable across the entire company all along, including full private conversations.

In an internal post, one employee wrote they were “incensed” that the scenario that some had warned about had come to pass.

Their post was titled “MCI Data: Not So Tightly Controlled” and screenshotted Bosworth’s earlier comment defending MCI’s security. “Hey Metamate, define the word ‘hubris’ for me,” another employee commented, referring to Meta’s internal AI assistant. “I swear we’re probably the richest yet most incompetent company to exist in tech right now,” another chimed in.

It told Business Insider that it suspended — but not outright canceled — the program and that, going forward, it would be optional.

Whether the data proved useful is another question.

In a July Q&A with executives, a Meta manager said it had not been used to train its latest AI models, according to another person familiar with the matter.

At a July town hall, Bosworth said the data was more useful than Meta thought it would be, according to two people familiar with the matter.

Meta declined to comment on how, and if, the data was used.

Bosworth has made his position clear: MCI opponents didn’t improve Meta. “They made the entire company worse,” he said — three times — in a leaked July audio recording published by the activist group More Perfect Union. “I think the people who put up that petition on MCI thought they were helping.

They weren’t helping; they made it a lot harder for us to make changes internally,” he said.

To the people who worked against MCI, the opposite is true.

Despite mass layoffs, they came together to push back against the harvesting of their data.

They say their activism forced Meta to make the program opt-in, if it’s ever reintroduced. “This would not have happened without brave acts of employee resistance,” the MCI Petition website reads.

After the intense backlash from Meta employees, other companies will think twice about forcing workers to give up their data to train them, said William Fitzgerald, the founder of advocacy and communications firm The Worker Agency. “It was a historic win,” he said. “Zuckerberg wanted more data — he tried to do it this way, but it didn’t work.” Have a tip?

Contact Charles via email at crollet@businessinsider.com or on Signal and WhatsApp at 628-282-2811.

Use a personal email address, a nonwork WiFi network, and a nonwork device; here’s our guide to sharing information securely.

Advertisement