Meta is leaning on its own employees to teach internal AI models how work gets done, even as it trims thousands of jobs. Mark Zuckerberg says the company gets better results when models learn by watching ”really smart people” handle real tasks, rather than relying on outsourced labor. Meta’s approach comes as the company cuts about 10% of its workforce, or roughly 7,800 people.
The pitch is simple enough: if you want a model to imitate good judgment, start with people Meta thinks have it. That framing also says plenty about where the company sees value inside its own walls. Big tech loves automation, but it still seems to trust human expertise most when the humans are already on payroll.
Why Meta prefers its own employees over contractors
Zuckerberg said the average skill level inside Meta is higher than what the company can typically get through outsourcing channels. That is both a compliment and a bit of a swipe at the contract workforce model many firms have relied on to label data, review output, and keep AI projects moving cheaply. Meta appears to be betting that its internal teams produce cleaner training signals and, just as important, less mediocre AI.
It is also a reminder that AI training is still painfully dependent on human work, even when the end goal is to reduce it. The industry has spent the past two years chasing larger models and more automation, but the bottleneck often remains the same: better inputs produce better systems.
Layoffs first, then a month of notice
Meta said in April that the cuts would affect about 10% of its workforce. Unlike many companies that move quickly once the axe is out, Meta gave staff almost a month’s notice, although it never published a specific list of who would go. That slower approach may have softened the shock a little, but it did nothing to make the optics kinder: the company is reducing headcount while asking remaining employees to help train the machines that may replace parts of their work.
The contradiction is hardly unique to Meta. Amazon, Google, and other giants have spent recent years talking up AI efficiency while simultaneously reorganizing teams around it. The difference here is the bluntness of Zuckerberg’s logic: the company wants the smartest available examples, and it thinks those examples are already in the building.
What Meta is really trying to optimize
At a practical level, this is about speed, quality, and control. Training with employees can give Meta better access to internal workflows, domain knowledge, and the kind of judgment that outsourced annotators may not have. It also keeps a more sensitive part of the AI pipeline closer to home, which matters when the product being trained is expected to reflect how the company itself works.
The larger question is whether that strategy becomes standard across big tech. If Meta can show that internal employees produce better models than cheaper external labor, others will copy it fast. If not, the company may end up paying premium salaries to build systems that still need a lot of outside help. That would be awkward, which is rarely enough to stop Silicon Valley.