Roughly 6,500 engineers and product managers were given a join or quit ultimatum and reassigned to generate puzzles and coding problems that train AI systems — work that Meta had already paid $14.3 billion to acquire a company built specifically to do.
An employee hijacked a livestreamed Meta all-hands this week, shouting an expletive demand at a senior AI executive while a presenter covered their face. The outburst was a scene, but the structural fact behind it is not: roughly 6,500 engineers and product managers were drafted into Meta's three-month-old Applied AI unit under a join-or-quit ultimatum, then reassigned from their jobs to generate puzzles and coding problems that train the AI systems their employer says will produce superintelligence. Three of them, speaking anonymously to WIRED, called it a "gulag."
The Applied AI unit, internally AAI, sits inside Meta Superintelligence Labs, the company division that chief AI officer Alexandr Wang runs after Meta acquired his former employer Scale AI for $14.3 billion. Scale built its business on exactly the work AAI is now producing: humans paid to write, rank, and evaluate the puzzles and coding tasks that frontier AI models learn from. Meta's stated reason for buying Scale was that the company had solved the human-data bottleneck. The current practice is to draft in-house engineers into the same work, because, as Mark Zuckerberg put it in a leaked internal audio justifying the reassignment, Meta employees are "significantly higher" in intelligence than outside contractors.
That justification is the hinge. The conscripts were not hired for the work. Some received a surprise email in early April and were told to join AAI or leave the company. Others were pulled in later batches under the same terms. The unit, which TechCrunch reports is led by Maher Saba, a 12-year Meta veteran who previously ran Reality Labs, was structured with up to 50 employees per manager. The reporting line runs up to chief technology officer Andrew Bosworth.
Inside AAI, the daily work is narrow. Engineers write or evaluate coding problems and puzzle-style exercises used to train and grade AI agents, the kind of model that completes multi-step tasks on a computer. Some have been asked to finish two such tasks per week. In WIRED's reporting, multiple staff described the output as a treadmill with no visible product, no public shipping target, and no obvious way to refuse. They used the word "gulag" because the framing inside the unit is an explicit internal claim that AAI is a "waypoint, not a destination," staffed by "very talented people" who will rotate out as soon as Meta builds another role for them. In the engineers' reading, the rotation promise is the relief valve that has not actually turned.
The conscription is not happening in isolation. Meta laid off about 8,000 employees, roughly 10 percent of its workforce, last month as part of a restructuring the company framed as an AI pivot. More than 1,600 current employees signed a petition protesting a separate U.S. click and keystroke monitoring program that the company has used to harvest AI training data from internal productivity tools, a program Meta has since partially walked back, allowing 30-minute pauses and exemptions. Across divisions, morale in the company's own internal language is at a record low. At an Instagram all-hands this week, chief product officer Chris Cox called the past months "brutal" and "the insanity of this company," then tried, per WIRED, to talk employees back from the maximalist position that AI will replace them and the minimalist position that AI is overhyped.
Mark Zuckerberg responded Friday with an internal memo acknowledging distress and mistakes and pledging no additional mass layoffs this year, a cap on manager-to-report ratios, bigger team-event budgets, a large hackathon next month, and a year-end desk reassignment. The memo, reported by WIRED, does not retreat from the AAI program or the join-or-quit framing that produced it. The 50-to-1 manager-to-report ratio is a problem Zuckerberg named. The function AAI is performing is a problem the memo treats as central, not contingent.
The reader's takeaway is not that Meta is uniquely bad at managing its AI workforce. The reader's takeaway is the labor playbook. Every frontier AI lab has, at some point in the last three years, treated high-quality human-generated data as the bottleneck on capability. Every frontier AI lab has, at some point, discovered that the cheapest way to produce that data at scale is to draft the people already inside the company. Meta is the first to acquire a $14.3 billion data-labeling company on the way to the same conclusion. That gap between the stated ambition and the labor practice is what carries forward into the next time the same story breaks about Google, OpenAI, or Anthropic. The "draftees" of Meta Superintelligence Labs are not a fringe complaint. They are the template.
What to watch: a public Meta statement on AAI and the join-or-quit dynamic, which the company has not made on the record. A follow-up from WIRED or other outlets on the identity of the executive named in the hijacker's demand; both reports decline to print it.