Meta reassigned thousands of engineers to label AI data, and the unit is in revolt

Gergely Orosz published a blunt question on June 16. His Pragmatic Engineer post asked why Meta is “destroying” its engineering organization, and it set off a week of debate. His reporting, in short: Meta has redirected a large share of its engineers away from product and toward generating AI training data, and a lot of them are furious.

That framing is sharp, and some of the specifics rest on insider accounts rather than anything Meta has confirmed on the record. So let’s separate what multiple outlets have reported, what comes from a single source, and what Meta itself says. The pattern underneath is the part that travels beyond one company: a profitable engineering org getting reshaped around AI faster than its own people can absorb.

What Orosz reported

The core claim is a reassignment, not a layoff. According to Orosz, Meta moved roughly 30 to 50 percent of engineers on core product, infrastructure, and security teams into an internal group he calls Agent Data Optimisation. The work there is data labeling and reinforcement learning from human feedback: writing coding puzzles that current models cannot solve, building the tests that check them, and grading AI-written code. Orosz estimates about 6,500 people sit in that org, of which 4,000 to 5,000 are software engineers, or roughly one in five or six engineers at the company.

He ties this to a broader cultural shift. For about two decades, he argues, Meta ran a high-performance engineering culture that treated software as a profit center. The reassignments, in his telling, recast a chunk of that workforce as a cost center almost overnight. He also points to second-order damage, including a claim that the Instagram Trust and Safety team lost around half its staff to labeling work and layoffs, which he links to a late-May security incident affecting high-profile accounts.

Treat the precise percentages as Orosz’s reporting rather than settled fact. Meta has not published a headcount breakdown of the unit, and the company did not confirm the figures publicly. What is independently corroborated is that the unit exists, that it’s large, and that the people inside it are unhappy.

The revolt other outlets confirm

The unhappiness is the most heavily reported part. TechCrunch reported on June 12 that engineers inside Meta’s roughly three-month-old Applied AI unit described being “drafted” into it with little real choice. One employee told reporters, “It’s literally the gulag.” Another said, “Most people find the work soul-crushing.” The unit, staffed by about 6,500 engineers and product managers, was assembled to manufacture coding problems and grade model output.

Then there is the monitoring. More than 1,600 Meta employees signed a petition protesting a program that tracks clicks and keystrokes to harvest AI training data, per TechCrunch’s reporting. The surveillance angle predates the June revolt. Back in April, Fortune reported that Meta planned to record employees’ screens and keystrokes to train its agent tools. A Meta spokesperson framed it plainly to Fortune: “If we’re building agents to help people complete everyday tasks using computers, our models need real examples of how people actually use them.” Reuters later reported, as Orosz notes, that Meta added limited controls letting staff pause the tracking for up to 30 minutes at a time and request exemptions.

This sits on top of real cuts. In May, Meta carried out an AI restructuring that eliminated about 8,000 roles, roughly 10 percent of its workforce, with more cuts floated for later in the year. The reassignment is a different lever than those layoffs. Where one removes people, the other repurposes them. Both point the same direction.

What Meta says back

Meta has not stayed silent, and a balanced read has to include its side. Chief product officer Chris Cox addressed what TechCrunch described as a “brutal” environment on a call with employees. Mark Zuckerberg sent an internal memo, described by TechCrunch, acknowledging that recent changes had “caused distress” and that the company had made mistakes it planned to address. According to follow-up reporting, Zuckerberg promised no further company-wide layoffs for the rest of 2026, moved to cap a manager-to-report ratio that had ballooned toward 50-to-one on some teams, and said the company would work to find new roles for engineers stuck in model-training work.

What the memo reportedly didn’t do is reverse the policy. There was no commitment to undo the join-or-quit transfer mechanism, restructure the unit’s mandate, or reconsider whether some of the company’s highest-paid engineers are the right people to be annotating data. The company’s broader position is that this is reorganization, not retreat: Meta’s revenue and free cash flow are both up, and the spending shift is toward AI infrastructure, not away from headcount costs it cannot afford. The friction, by that reading, is the cost of moving fast on AI rather than a sign the business is in trouble.

This is also not Meta’s first AI-unit shake-up. CNBC reported in October that Meta cut about 600 roles from its AI organization as Alexandr Wang, hired from Scale AI as chief AI officer, consolidated control over Meta Superintelligence Labs. The current turmoil is the latest chapter in a reorganization that has been running for the better part of a year.

Why this is the signal

Strip away the Meta-specific drama and the interesting part is the precedent. A company with thousands of senior engineers decided the highest-value use of a large slice of them, for now, is producing training data that no existing model or outside contractor can generate. That is a bet that human engineering judgment is more valuable as model fuel than as shipped code. Whether that bet pays off is genuinely unknown, and the people making it haven’t shown their math.

The morale story is the leading indicator worth watching. Orosz cited a reported jump in Meta engineers signing up for interview prep on interviewing.io in May versus a year earlier, a soft signal of attrition pressure rather than a hard number. He also reported that some engineers had their performance measured partly by AI token usage, the kind of metric that rewards looking busy with a model over shipping something that works. That’s the quiet risk in a lot of these AI mandates: the measurement changes before the work does, and people optimize for the new dashboard. If the most capable engineers can leave, surveillance and “soul-crushing” annotation work give them a reason to. The danger for any company running this playbook is not the labeling itself. It is losing the people who could have built the next thing while they are busy grading the last one.

What this means for you

If you work in or near a large engineering org, this is the case study to keep an eye on, not a panic signal. The useful takeaway is narrow: watch whether “AI builder” style retitling at your own employer comes with a real change in what you ship, or just a change in what gets measured. Meta’s engineers reportedly found their work recast around AI token usage and labeling throughput, and the backlash followed the metrics. The thing to track over the next two quarters is retention, specifically whether senior infrastructure and security people stay. That number, more than any memo, will tell you whether this restructuring worked or whether Meta spent its best engineers on a detour. As for the specifics still sourced to insiders, the exact percentages and the unit’s long-term mandate, treat them as reported until Meta confirms them.