Meta has officially scrapped AI tool consumption from engineer performance reviews. According to an internal memo obtained by The Information, executives Maher Saba and Santosh Janardhan informed staff that vanity AI dashboards and token counters will no longer play a role in performance evaluations. Moving forward, engineering leadership is pivoting back to fundamentals: code quality, delivery speed, and technical complexity.
The reversal highlights the predictable failure of forcing AI adoption through blunt administrative quotas. When leadership tied career progression to AI usage, engineers responded with "tokenmaxxing"—needlessly burning massive inference workloads simply to climb internal leaderboards. That artificial demand threatened to balloon Meta's internal compute spend into billions of dollars by 2026, forcing executives to outline centralized budgeting controls starting in 2027.
At the same time, Meta’s rollout of Hatch—an internal autonomous agent designed to execute desktop tasks—has run into employee pushback over account access and privacy, as reported by WIRED.
Goodhart’s Law remains undefeated: when compute consumption becomes a target, developers will automate waste rather than value. For engineering leaders rushing to mandate generative AI metrics, Meta's retreat is a blunt case study—tying performance to raw token volume manufactures enterprise-scale expense, not productivity.