Meta Engineers Call AI Unit a ‘Gulag’ as Zuckerberg Admits Mistakes

Meta built a new Applied AI unit three months ago and packed it with roughly 6,500 engineers and product managers. Now many of those workers want out. A Wired report describes the unit as near open revolt, with workers calling themselves "draftees" and one telling the outlet the work is "literally the gulag."

The breaking point arrived this week when someone hijacked a livestreamed, employee‑only presentation and launched into an expletive‑laden meltdown, demanding that the audience call a senior Meta AI executive "a piece of sh*t." One presenter reportedly covered their face with both hands.

Workers say Meta gave them no real option: join the unit or quit. According to the Wired account, many discovered the reassignment through a sudden email. One employee called the selection process "quite random" on Reddit, as Business Insider previously reported. An internal announcement explained the reasoning: Meta's models still could not beat humans at technical work like coding, and the company needed real examples of how people complete everyday tasks.

Their daily task is narrow and repetitive: writing puzzles and coding problems used to train Meta's models. "Most people find the work soul‑crushing," one employee told Wired. Another described the environment bluntly: "It's literally the gulag."

A leaked audio recording captured Zuckerberg defending the decision to draft staff instead of hiring outside contractors. He pointed to his chief AI officer, Alexandr Wang, who sold his data‑labeling startup Scale AI to Meta for $14.3 billion before taking over Meta Superintelligence Labs, per the Wired report. The average Meta employee, Zuckerberg argued, carries "significantly higher" intelligence than a third‑party contractor — which, in his view, made staff the better pick.

By Friday, Zuckerberg was in damage‑control mode. In an internal memo, he admitted the recent changes had "caused distress," according to the Wired report and owned the missteps. "Given the complexity of these changes, we've made mistakes and will almost certainly make more," he wrote, per Wired. He also ruled out further company‑wide layoffs for the rest of 2026.

The Applied AI unit is led by Maher Saba, a 12‑year Meta veteran who once served as a vice president in Reality Labs — the division that burned through $83 billion on the metaverse before the company pivoted to AI, according to the Wired account. The new group reports up to Meta's chief technology officer, Andrew Bosworth.

At the start, the structure was punishing: up to 50 employees answered to a single manager. The discontent spread beyond the Applied AI group. More than 1,600 employees across the company have reportedly signed a petition against a program that tracks their clicks and keystrokes to gather AI training data.

Zuckerberg's Friday memo came with concrete offerings. Meta promised bigger budgets for team offsites, a company‑wide hackathon in July, the return of assigned desks in many offices, and fewer reports per manager. He also pledged to find new roles for staff stuck doing model‑training work.

The May overhaul that created the Applied AI unit cut roughly 8,000 jobs and shifted 7,000 more people into AI roles, touching close to a fifth of Meta's total headcount. The company is far from alone in turning itself inside out for AI, but the revolt inside its walls shows how quickly a top‑down AI transformation can breed resentment when engineers feel like raw material rather than talent.

For the broader tech workforce, Meta's Applied AI revolt is a warning about what happens when companies treat highly skilled engineers as interchangeable data‑labeling resources. The model‑training work being done by Meta's drafted engineers is the same kind of task that companies like Scale AI built entire businesses around — except Meta is using $300K+ engineers instead of contract workers.

The fallout also raises questions about the quality of that training data. If the people writing the puzzles and coding problems are resentful and looking for exits, what does that mean for the models being trained on their output? As AI companies race to build the next generation of models, the human cost of that training pipeline is becoming impossible to ignore.