Inside Meta's 'Soul-Crushing Gulag': 6,500 Engineers Revolt Against Forced AI Data Labeling

When Zuckerberg told his employees last March that Meta was forming an Applied AI unit and that he needed some of the company’s best engineers to staff it, the framing was aspirational: this was a chance to be at the center of Meta’s most important strategic initiative, building the data infrastructure for a future superintelligent system. Three months later, the engineers who were placed into that unit — roughly 6,500 of them, many without meaningful choice — are describing their experience in strikingly different terms.

“It’s literally the gulag,” one engineer told TechCrunch. “Most people find the work soul-crushing,” said another. A third described themselves as a “draftee” — someone ordered to a posting they did not seek and cannot easily leave. The unit’s primary function, as these employees tell it, is generating puzzles, coding challenges, and synthetic scenarios that will be used to train Meta’s next generation of AI models. The work is repetitive, unglamorous, and, for people hired to build products or conduct research, deeply demoralizing.

The Event That Broke Containment

For months, the frustrations inside Meta’s Applied AI team circulated on anonymous forums like Blind and in private Slack channels. What changed on Thursday was that someone brought the fight into the open — literally. During an all-hands livestream presentation for the team, an unknown individual hijacked the broadcast with an expletive-laden outburst, demanding that the audience pass along a profane message to a named senior Meta AI executive. One of the presenters was visibly shaken and covered their face.

The clip spread within the company and outside it within hours. The incident was notable not just for its rawness but because it happened in a managed internal forum, the kind of space where dissent is usually carefully modulated. The fact that someone felt compelled — or emboldened — to do it in that setting signals a level of collective frustration that normal internal channels have failed to contain.

The Petition and the Surveillance Grievance

The livestream incident did not occur in a vacuum. More than 1,600 Meta employees across the company — not just those in the Applied AI unit — have signed an internal petition protesting a separate but related policy: the monitoring of their mouse movements, application usage, and keystroke patterns for the purpose of AI training data collection. The petition argues that using employees as unwitting subjects of behavioral data harvesting, even with broad consent clauses buried in employment agreements, is ethically wrong and undermines the trust relationship between employer and employee.

This grievance reflects a broader tension in how technology companies are approaching the challenge of generating high-quality AI training data. Human-generated data — particularly data produced by skilled knowledge workers performing cognitively demanding tasks — is now viewed as a premium resource for training advanced AI models. But extracting that data from employees, without meaningful consent or compensation, is a form of labor extraction that is drawing increasing pushback.

The Structure That Created This Problem

The Applied AI team was established in March 2026 as a support function for Meta Superintelligence Labs, the division led by Alexandr Wang — the founder and former CEO of Scale AI — whom Zuckerberg hired as Meta’s chief AI officer earlier this year. Meta Superintelligence Labs is the high-prestige research division charged with pursuing AGI. The Applied AI team is, in effect, its support staff.

The organizational design created an almost textbook formula for resentment. Engineers who had built careers as autonomous individual contributors or led their own teams were placed under a unit structure in which, initially, up to 50 engineers reported to a single manager — a ratio that made meaningful supervision or career development virtually impossible. The unit is led by Maher Saba, a 12-year Meta veteran previously based in Reality Labs, who reports to CTO Andrew Bosworth.

Meta assigned engineers to the unit through a process that employees described as having a binary choice: accept the assignment or leave the company. There was no competitive application process, no opt-in mechanism, and little transparency about how specific individuals were selected.

Zuckerberg’s Response

Mark Zuckerberg acknowledged the situation in an internal memo this week, admitting that recent organizational changes had “caused distress” and promising to address “identified mistakes.” The memo did not signal any reversal of the core strategy — using in-house engineers to generate AI training data — but acknowledged that the implementation had been “disruptive” in ways the company had not fully anticipated.

In leaked audio from a separate internal discussion, Zuckerberg justified the use of engineers for data-labeling work by arguing that Meta employees possess “significantly higher” cognitive abilities than third-party contractors, making them preferable for generating the high-quality synthetic data needed to train frontier models. The argument, while coherent from a data quality standpoint, did nothing to address the employees’ core complaint: that they were not hired to be data labelers, and that this deployment represents a fundamental breach of the implicit employment contract.

The Larger Stakes

What is happening inside Meta’s Applied AI unit is, in miniature, a preview of a question that will define the relationship between knowledge workers and AI companies for the coming decade. As AI models become central to every major company’s strategy, the demand for high-quality training data produced by skilled humans will only grow. The question is who bears the cost of generating that data — and what happens when companies decide the most efficient answer is: their own employees.

For Meta specifically, the employee revolt comes at a precarious moment. The company is betting enormous sums — $65 billion in 2026 capital expenditure alone — on its AI strategy, and Superintelligence Labs is the crown jewel of that bet. If the Applied AI team cannot generate the training data that Superintelligence Labs needs, or if the internal dysfunction disrupts the data pipeline, the downstream impact on Meta’s model quality could be significant.

The situation also risks becoming a talent retention problem. Meta’s compensation has historically been one of its key tools for attracting top engineers. If those engineers feel they have been conscripted into work that diminishes their skills and autonomy, some will leave — and in a market where top AI talent is being aggressively recruited by every major player, those departures will be difficult to replace.