Amazon workers under pressure to up their AI usage are making up tasks
Corporate pushes to increase employee use of generative AI—often tracked via token consumption and internal leaderboards—are leading engineers to fabricate tasks and run pointless agents just to keep their metrics green. Commenters link this to Goodhart’s law and perverse incentives, comparing token burn to lines-of-code counting, travel-spend quotas, or Soviet-style production targets that reward waste over outcomes. Many see real experimentation value in AI tools, but argue that tying performance or budgets to raw usage distorts behavior, obscures true productivity, and has non-trivial financial and environmental costs.
Perverse incentives & Goodhart’s law
- Many see token-usage targets as a textbook case of Goodhart’s law: once “tokens consumed” becomes a goal, it stops reflecting real productivity.
- Comparisons are made to past bad metrics like lines of code, bug bounties, and LOC-based performance reviews that encouraged wasteful behavior.
- Some argue leadership mainly wants a green dashboard and “AI adoption” numbers to justify big AI investments or please investors.
How employees game token metrics
- People describe using AI for low-value work: auto‑docs, unit tests for everything, endless diagrams, or agents that churn nonsense and delete outputs.
- Internal leaderboards and implied links to performance reviews reportedly trigger a race to “tokenmaxx,” even at FAANGs and large enterprises.
- Some joke about tools specifically created to burn tokens, or chaining agents to maximize usage.
Debate on real productivity vs. busywork
- Supporters say forcing everyone to try AI accelerates discovery of genuine use cases; experimentation necessarily includes waste.
- Detractors argue trivial tasks done via AI are slower and far more expensive than known commands, scripts, or linters, especially when results must be reviewed for hallucinations.
- Some report modest or unclear productivity gains (e.g., slight PR/velocity increases despite huge spend); others claim dramatic speedups on their own teams.
- There’s disagreement over whether AI lets people “do things without knowing things” (a positive abstraction) or dangerously erodes core skills.
Management culture, fear, and metrics
- Commenters describe executive pressure, AI trainings/hackathons, and slogans like “AI revolution/era,” often feeling coercive and optics‑driven.
- Several note that anxious engineers, influenced by social media stories of AI-native orgs and firings, burn tokens to avoid being labeled laggards.
- Some compare the whole situation to RTO mandates, DEI fads, or Soviet-style central planning: top‑down quotas, dashboards, and box‑ticking.
Environmental and economic concerns
- Multiple posts criticize burning compute “for nothing” during a climate crisis, linking token quotas to data center expansion and energy use.
- Others highlight circular financial incentives: big firms invested in AI providers are effectively paying themselves by driving internal usage, with little clear ROI.
Overall sentiment
- Strong skepticism dominates, but a minority see structured “overuse” as a necessary, if clumsy, way to learn how AI can genuinely help.