Meta's months-old AI unit is a soul-crushing gulag, say the engineers stuck inside it

Inside Meta's three-month-old Applied AI team, employees call themselves "draftees." As in: they were drafted into a unit they did not choose, assigned work they find soul-crushing, and given the option to either accept it or leave the company.

The details, reported by Wired this week, paint a picture that is at odds with every polished recruiting pitch about the future of AI. The unit of 6,500 engineers and product managers was created by reassigning existing Meta employees — many of whom were working on projects they had chosen — into an organization tasked with generating puzzles and coding problems to train AI models. Up to 50 employees were initially assigned to a single manager. An internal presentation was hijacked by someone who called a senior Meta AI executive a profanity. More than 1,600 Meta employees across the company have signed a petition protesting a program that monitors their clicks and keystrokes for AI training data.

Meta's chief product officer, Chris Cox, called the current environment "brutal" on an internal call. CEO Mark Zuckerberg acknowledged in a memo that recent changes had "caused distress" and admitted the company had made mistakes it plans to address.

The memo and the Wired report together reveal something that the AI industry's growth narrative rarely acknowledges: the human cost of training artificial intelligence is not just measured in electricity and compute. It is measured in the working conditions of the people who produce the training data.

The Labor Model Behind the AI Revolution

The Applied AI team's work — generating puzzles, coding problems, and structured tasks that AI models learn from — is essential to building better AI. But the way Meta organized this work tells you something about how the company views the humans who produce it.

Rather than hiring people specifically for this role, Meta reassigned existing employees from other teams. The unit is led by Maher Saba, who previously oversaw a division in Reality Labs — the metaverse unit that burned through $83 billion before Meta pivoted to AI. The reporting structure initially had up to 50 people per manager, a ratio that makes individual career development impossible and creates a production-line atmosphere.

The word "gulag" that employees use to describe their experience is hyperbolic but revealing. It describes a system where people are assigned to work they did not choose, managed at scale without individual attention, and given the implicit message that their role is to serve the machine rather than develop their careers. For engineers who joined Meta to build products, being reassigned to generate training data for AI models is a demotion — even if the company frames it as a strategic priority.

This is not unique to Meta. Every major AI company employs thousands of people — some directly, many through contractors — to produce the data that trains their models. Anthropic, OpenAI, Google, and Microsoft all use similar workflows. The difference is that most of this work is done by outsourced annotators in low-wage countries, invisible to the companies' public narratives. Meta's mistake was pulling this work in-house and making it visible to its own highly compensated engineers, who have the leverage and the platforms to complain about it.

The Keystroke Monitoring Problem

The petition signed by 1,600 employees protesting Meta's keystroke monitoring program is about more than privacy. It is about the relationship between a company and its workforce when the company's primary product is built on data extracted from that workforce.

Meta's monitoring program, which tracks employee clicks and keystrokes to capture data for AI training, is the logical endpoint of a labor model that treats human activity as raw material for machine learning. When your work product is data — when every click, every edit, every problem you solve becomes a training example for an AI model — the boundary between work and surveillance becomes meaningless. The company is not just monitoring your productivity. It is harvesting your cognitive output to train the systems that may eventually replace aspects of your job.

The engineers' objection is not irrational. They can see the trajectory: their keystrokes train AI models, the models get better at generating code and solving problems, and eventually the models require fewer human keystrokes to produce the same output. The question of whether AI will replace software engineers is debated endlessly. The question of whether AI will reduce the number of engineers needed for a given amount of work is not debated at all — it is already happening.

The $83 Billion Precedent

It is worth noting that the Applied AI team is led by a manager who previously oversaw Reality Labs, the division that represents Meta's most expensive pivot. The metaverse burned $83 billion before Meta moved on to AI. The company's track record with massive strategic bets that require enormous human and financial capital is, charitably, mixed.

The concern for Meta employees — and for anyone watching the company's AI strategy — is whether the Applied AI team's current structure represents a sustainable investment in a core capability or whether it is another $83 billion-style bet that will be abandoned when the next strategic priority arrives. Employees who were reassigned from other teams have reason to wonder whether they will be reassigned again in 18 months, this time into whatever comes after AI.

What This Means For You

If you work in tech, the Meta Applied AI story is a reminder that the AI revolution is not just changing what products get built. It is changing how the people who build those products are managed, monitored, and deployed. The "gulag" language is extreme, but the underlying frustration is real: when companies treat human talent as interchangeable inputs for a data pipeline, the talent eventually pushes back.

If you are an investor in AI companies, the question is whether Meta's labor model is an outlier or a leading indicator. The answer is probably both. Meta is worse than most at managing its workforce through pivots, but every AI company faces the same structural problem: training data is expensive to produce, the humans who produce it have more leverage than the companies would like, and the monitoring required to capture that data creates legitimate employee grievances.

For everyone else, the Meta story is a window into the actual mechanics of AI development. The polished demos and the trillion-dollar valuations depend on thousands of people generating puzzles and coding problems in a structure they describe as soul-crushing. The AI that writes your emails and generates your images and summarizes your meetings was trained on the cognitive labor of real people working under conditions they find degrading. That does not make the technology less useful, but it should make the narrative about AI as a frictionless, autonomous system more honest about the human inputs that make it work.