Meta's new AI division staff discontent erupts: accusations liken it to living in a concentration camp with souls crushed, engineers suffer immensely

Meta forces 6,500 engineers to work on AI data labeling, employee dissatisfaction gradually erupts, reflecting the hidden cost of the current AI arms race being rarely discussed publicly.
(Background: Meta scandal exposed: Zuckerberg demands monitoring employees' keyboard and mouse recordings, claiming "it does not affect performance evaluations")
(Additional context: Training AI with iPhone-wrapped heads: cheap laborers are becoming robot flesh teachers, teaching it to carry goods, work, and do housework)

Table of Contents

Toggle

• Engineers drafted by an email
• The calculator bought with 14.3 billion dollars
• Data is the real bottleneck

A Meta employee live presentation went out of control this week: someone broke down emotionally, swearing, and asked the audience to tell a senior Meta AI executive that he is "a piece of sh*t." What is happening at one of the world's most valuable tech companies?

Engineers drafted by an email

According to Wired, Meta’s "Applied AI" team was formed just about three months ago and is already facing internal resistance. This new department, with a scale of 6,500 engineers and product managers, many employees only learned they were transferred into Applied AI after receiving "a sudden email." A self-described draftee on Reddit described the entire process as "quite random" and completely unanticipated.

Business Insider’s internal memo explains the reasons for the transfer: Meta’s AI models still lack the knowledge to outperform humans in technical tasks like programming. Simply put, the models are not smart enough and require humans to create examples manually.

These engineers are tasked with generating puzzles and programming problems to train AI models. One employee told Wired: "This is basically a Gulag," and another said, "Most people feel this work is soul-crushing."

The calculator bought with 14.3 billion dollars

In a leaked internal meeting recording, CEO Zuckerberg explained why external contractors are not used. His reasoning has two layers:

• First, Alexandr Wang, who sold the data annotation startup Scale AI to Meta for $14.3 billion, then became Meta’s Chief AI Officer and led Meta Superintelligence Labs, is very familiar with data labeling.
• Second, Zuckerberg stated that Meta’s regular employees have "significantly higher IQ" than third-party contractors, making them a better choice.

This logic sounds commercially rational but overlooks a key premise: are high-paid engineers willing to accept this arrangement?

Data annotation is the core of AI training. Simply put, it involves humans using their judgment and demonstrations to teach AI "what good answers look like." When a model underperforms in complex reasoning or programming tasks, the problem is often not the algorithm but the lack of high-quality human demonstration data during training.

Meta previously acquired Scale AI for $14.3 billion, aiming to "enable high-quality human annotation." But forcing 6,500 people to do this without choice, one employee described feeling "there was no real choice: join or leave."

Data is the real bottleneck

On the surface, this is a workplace management failure; deeper down, it’s a declaration in the AI race: data quality is the true limit of current large model capabilities.

Companies spend on computing power, publish papers, and compete over parameters, making the competition seem like an engineering problem. But Meta’s actions reveal another aspect: when models underperform in real tasks, the bottleneck for engineers isn’t architecture but the quality and scale of demonstration data. Using the smartest people for the most tedious annotation work itself indicates that the large model competition has entered a new stage.

OpenAI, Google, and Anthropic are also building data, just differently and mostly privately. Meta’s case exploded because it forced and internalized this process, leaving no way out for employees.