It started with a surprise email in late April 2026. Engineers across Meta — from infrastructure to security teams — received notifications that they were being moved to a new unit called Applied AI, internally known as Agent Data Optimization or ADO. They didn't apply. They weren't promoted. It wasn't optional.
Employees started calling themselves "draftees."
The unit, reportedly 6,500 people strong with about 4,000 to 5,000 software engineers, has one mandate: produce training data for Meta's AI models. Workers write coding problems. They write tests. They grade AI-generated code. They generate step-by-step puzzle solutions that models learn from. It's the kind of work usually outsourced to third-party data-labeling contractors.
"It's literally the gulag," one employee told Wired, describing the day-to-day reality of the unit.
"Most people find the work soul-crushing," another employee said of the repetitive nature of generating training puzzles and coding problems.
The frustration runs deep because these are experienced engineers — people who joined Meta to build products, solve hard problems, and ship code. Instead, they're doing work they describe as beneath their skill level, with ultra-flat management structures where up to 50 people report to a single manager. Career development has effectively stopped.
Zuckerberg's defense of the program made things worse. In leaked internal audio, he explained that Meta uses its own engineers instead of contractors because Meta employees have "significantly higher" intelligence. The framing landed badly with workers who were conscripted without consent.
The anger has spilled into public view. An employee hijacked an internal presentation's live audio with an expletive-laden outburst directed at a senior AI executive. One presenter reportedly covered their face with their hands.
More than 1,600 employees company-wide have signed a petition protesting a related program that monitors workers' clicks and keystrokes for AI training data — a practice employees describe as invasive surveillance disguised as productivity measurement. The petition is called the Model Capability Initiative.
Meta's chief product officer Chris Cox was compelled to address the "brutal" work environment directly on a company-wide call. Zuckerberg eventually released a Friday memo admitting the company "made mistakes" and that the changes "caused distress."
The toll is visible in attrition projections. Analysis suggests roughly 1,900 of the 6,500 ADO engineers will leave voluntarily by the end of 2026 — about 3.8 times Meta's normal baseline attrition rate. The strongest performers are leaving first, the ones with the most leverage and the cleanest recent-project narratives.
For workers still stuck in ADO, the question isn't whether to leave — it's when. And for Meta, the experiment in converting its own engineering workforce into training data generators is proving that quality of output depends heavily on the willingness of the people producing it.