Meta Reassigns Thousands of Engineers to AI Data Labeling Roles

Meta has involuntarily reassigned thousands of software engineers from core product teams into a new internal unit focused on artificial intelligence training data. The Agent Data Optimization group, known internally as ADO, now employs roughly 6,500 people, with an estimated 4,000 to 5,000 of them being software engineers. The move represents one of the largest internal workforce redirections in Silicon Valley history.

Engineers Turned Into Data Labelers

Starting in late April, product engineering teams received mandates to transfer 30 to 50 percent of their staff into ADO. The reassigned engineers now spend their days performing reinforcement learning from human feedback tasks, including writing test suites for AI-generated code, packaging work into Docker containers, and reviewing model outputs across multiple iterations. One affected engineer described the experience to the Pragmatic Engineer newsletter as having zero purpose. Infrastructure and security teams were hit particularly hard, with Instagram's Trust and Safety team losing roughly half its staff.

Why Meta Made the Shift

The reorganization reflects Meta's push to build large language models competitive with OpenAI and Anthropic. Training frontier AI models requires massive volumes of high-quality labeled data, and Meta concluded that its own engineers produce better results than external contractors. By redirecting engineering talent toward data generation and model refinement, Meta is betting that in-house expertise will yield superior training data.

Fallout Across the Company

The consequences have been swift. A June 1 security breach led to account takeovers affecting high-profile Instagram users, followed by the resignation of Meta's chief information security officer Guy Rosen on June 2. A platform-wide outage hit on June 12. Internally, morale has cratered, with job search platform signups from Meta employees surging through May and June.

Meta's leadership has acknowledged the problems. CTO Andrew Bosworth called the handling atrocious and promised better communication. CEO Mark Zuckerberg sent an internal memo conceding that mistakes were made and pledging to find improved roles for affected staff. The episode raises questions about whether converting experienced product engineers into data labelers is sustainable, or whether the loss of institutional knowledge will outweigh any gains in training data quality.