Amazon employees, pressured to increase AI usage metrics, are fabricating or padding tasks to hit token consumption targets.
Key Takeaways
Token usage has become a tracked performance signal at Amazon, with internal leaderboards and implied pressure even when officially denied.
Workers are generating artificial workloads to rank on token dashboards rather than applying AI where it adds value.
The pattern mirrors Goodhart’s Law: once token spend becomes a metric, it stops measuring actual AI adoption.
Kiro, Amazon’s internal AI coding tool, is central to the gamification – teams not directly billed for token use have little friction inflating numbers.
The dynamic extends beyond Amazon; multiple FAANG companies have similar dashboards and manager-level nudges toward higher token consumption.
Hacker News Comment Review
Commenters broadly agree the pressure is real but officially deniable – leadership disclaims that token spend affects reviews while simultaneously showing token dashboards in manager meetings.
A recurring analogy is Soviet-style quota gaming: workers running AI agents 24/7 on nonsense to top leaderboards, with one org reportedly running a continuous token-burning process called “GasTown.”
Several engineers note that Cursor inflates AI-attribution stats by claiming linter-level autocorrections as AI edits, suggesting the metric-gaming problem is tool-level, not just human behavior.
Notable Comments
@eunoia: Confirms FAANG leadership explicitly reviews token dashboards in weekly manager meetings while simultaneously acknowledging the metric is gameable.
@dtnewman: Built and open-sourced “Burn, Baby, Burn,” a CLI tool satirizing the pressure – signals how absurd the incentive has become to builders.
