The email subject line read “Applied AI — You Have Been Selected.” There was no opt-out. Within weeks, thousands of Meta engineers and product managers had been reassigned, by fiat, to the company’s newest artificial-intelligence effort.
The unit, called Applied AI, was formed three months ago. Roughly 6,500 engineers and product managers were brought in through what employees call “raid emails” — mass assignments with no choice and no discussion, according to people familiar with the matter. Their assignment: generate puzzles, programming problems and test questions to train Meta’s AI models.
The work has produced an unusual degree of public dissent inside a company known for tight control. More than 1,600 employees have signed a petition protesting the program, people familiar with the matter said. The petition objects to monitoring — keystrokes and clicks tracked for AI training — and to the way the unit was staffed.
Tensions spilled into the open this week when an internal livestream was hijacked. Employees used the chat to berate executives, and some called the environment “soul-crushing,” a phrase that has since spread through internal channels and, from there, into public view. Employees have begun calling the unit “the gulag,” a label that circulated widely enough that executives are aware of it.
Chris Cox, Meta’s chief product officer, acknowledged the problem in an internal session, calling the environment “brutal” and saying management had moved too fast. On Friday, Mark Zuckerberg sent a companywide note. He conceded the program had “caused pain” and promised changes, without specifying what they would be, people who read the note said.
The episode exposes a tension at the center of Meta’s AI push. The company is spending tens of billions of dollars a year on models, data centers and researchers, and it has said repeatedly that AI is its top priority. Applied AI was designed to produce the vast quantities of training data — problems, answers, explanations — that frontier models require. Meta has been explicit about wanting to reduce its dependence on outside data sources, including licensed content and public web scrapes.
What employees object to is not the goal but the method. Engineers hired to build products say they were drafted into data-production work without warning, that the assignments use their professional judgment in a narrow way, and that the monitoring treats them as sensors rather than staff. Several said in internal threads that the program asks experienced engineers to do work that could be done by contractors, and that the company has not explained how long the assignments will last.
Meta declined to comment on the petition. Mr. Cox told employees that the company is reviewing the program’s design and that changes are coming, without offering a timeline.
The public dimension is what makes this episode different from the usual internal grumbling. Meta is competing with OpenAI, Anthropic and Google for a small pool of AI talent, and engineers have options. Recruiters at rival labs said they have seen an uptick in outreach from Meta employees in the past week, though it is too early to say whether the petition will translate into departures.
The episode arrives at a delicate moment for Meta’s public positioning. The company has presented itself as a builder of open AI, releasing models that rivals can inspect and modify, and it has argued that openness gives it a talent advantage. The petition complicates that story by drawing attention to the human machinery behind the models, thousands of engineers generating training data under monitoring systems they did not choose. Employees close to the petition organizers said the strongest sentiment is not opposition to AI work but opposition to conscription. Several said they would gladly volunteer for data-generation projects; what they object to is being drafted, and watched, without a say in how long the assignment lasts.
The broader question is how far companies will go to secure training data. Meta’s approach — pulling thousands of employees into data generation — is one answer. OpenAI and Anthropic have relied more on contractors and on user interactions. Google has used employee feedback programs with voluntary participation. Analysts said Meta’s scale and urgency pushed it toward the most forceful option.
For Mr. Zuckerberg, the challenge is managerial as much as technical. He has bet the company’s next phase of growth on AI, and the Applied AI program is one of the most visible expressions of that bet inside the company. A revolt among the very engineers asked to build it does not change the investment plans, but it complicates the execution.
Employees said they are waiting to see what the promised changes look like. The petition’s organizers have asked for three things: an end to the monitoring, a choice about participation, and a public commitment on how long assignments will last. None of the three has been granted, people familiar with the matter said.
The company said it will respond this month. In the meantime, the puzzle-generation teams keep producing problems, the monitors keep tracking keystrokes, and the internal chat remains active. For a company that measures itself in engagement, the episode is being watched inside Meta as a test of whether the workforce’s complaints move the product roadmap — or simply get absorbed by it.
