Meta Drafted Thousands of Engineers Into AI Training. Six Weeks Later, It Is Letting Them Walk.

Six weeks after conscripting its own engineers, Meta is not so quietly handing the choice back. The company is reversing last month's decision to redeploy thousands of staff into AI training roles, telling them in an internal memo that it will now "defer to each individual's choice" about whether to help, according to a memo first reported by Business Insider. 

The reassignment, which employees had taken to calling "the draft," moved roughly 6,500 engineers and product managers into a new Applied AI unit built to generate the coding puzzles and challenge problems Meta's models could not yet solve. Most learned of the move by surprise email. The offer was simple: join the unit or leave the company.

A Meta spokesperson framed the new policy as giving employees more flexibility. The timing tells a different story.

The revolt that forced the reversal

The draft did not land well. Engineers being paid frontier-research compensation found themselves producing what amounts to expert-grade data annotation, the same category of work that contractors at firms like Scale AI handle for a fraction of the price. One employee told Wired the unit was "literally the gulag." By June, TechCrunch reported the division was on the verge of revolt, with workers describing the assignment as soul-crushing.

The backlash hit at a low point for the company's mood. CTO Andrew Bosworth told staff on a recent internal call that morale was "probably one of the worst it's ever been," and in a follow-up memo he conceded Meta had done an "atrocious job" explaining the vision behind the unit to the people it had moved into it. 

A morale fix that leaves the real problem in place

Letting engineers opt out solves the optics. It does not solve the constraint that created the draft in the first place. Synthetic training data, which drove the model gains of the last few years, has hit a quality ceiling at the frontier. To push a model past human performance on hard coding tasks, you need fresh human-written examples difficult enough that current models fail them and novel enough to leave no trace online. Meta decided its own engineers were the best available source of that data. Handing them an exit does nothing to refill the well.

This is the second whiplash from Meta's AI org in two months. We covered the first when the company moved to cap employee AI usage after the "tokenmaxxing" era spun out of control, an about-face for a company that spent the prior year ordering staff to use AI as aggressively as possible. The pattern is a leadership team writing the operating manual for an AI-first company in real time, then tearing out pages when the workforce pushes back.

Why it matters

Meta is the largest live experiment in rebuilding a giant company around AI, which makes its mistakes cheap to learn from and expensive to repeat. The lesson here is about people, not strategy. The logic behind the draft was sound. The execution treated highly paid specialists as a pipeline input, and people do not experience the strategy memo, they experience the task in front of them. Meta paid for that distinction with a reversal and a morale floor it now has to climb back from.

The brand bill comes next. Our colleagues at State of Brand have been tracking the gap between Meta's outward AI story and the reality inside the building, from automating advertisers' campaigns to the broader trap of branding AI as a "coworker" while your actual coworkers are in open revolt. A recruiting pitch that promised frontier research and delivered annotation work is the kind of gap that trails a company into every offer letter it sends for years.

The business itself is still booming and the AI bet is real. What this month showed is that you cannot draft your way to a culture. The reversal is Meta admitting it tried.