Meta Drops AI Usage Metrics from Engineer Reviews After "Tokenmaxxing" Gaming
Original title:Meta drops AI usage from engineer performance reviews after "tokenmaxxing" backfires
Meta has removed AI tool usage metrics from software engineer performance evaluations. The company had originally incentivized adoption by tying prompt activity and LLM engagement to reviews, but the policy quickly backfired into "tokenmaxxing"—engineers artificially inflating token consumption to satisfy quantitative performance targets. Meta's retreat illustrates Goodhart's law in enterprise generative AI: when tool usage itself becomes the performance metric, genuine engineering output is easily overshadowed by performative computation.
Why it's worth reading
A concrete lesson in Goodhart's law for the generative AI era, showing the operational hazards of mistaking tool utilization metrics for genuine engineering output.