SeattleAmazon workers invented tasks to satisfy mandatory AI-use targets.
Employees fed internal copilots fabricated work to hit usage quotas, decoupling measured adoption from productivity.
Amazon has not confirmed the scale or announced changes to how it tracks AI returns.
Sources: Hacker News
THE LAW
When a measure becomes a target, it ceases to be a good measure. Economist Charles Goodhart observed this in 1975 watching the Bank of England target money supply aggregates — the moment policymakers picked one, banks restructured around it and the aggregate stopped tracking what it used to track.
THE SOVIET PRECEDENT
Soviet nail factories famously gamed output quotas: when measured by weight, they produced a few enormous nails; when measured by count, millions of tiny useless ones. The lesson is that workers optimize the proxy, not the underlying goal — and the proxy always diverges.
THE COBRA EFFECT
British colonial Delhi paid a bounty per dead cobra. Residents began breeding cobras for the bounty. When officials cancelled the program, breeders released their now-worthless snakes, and the wild population exploded above its starting point. Every incentive scheme creates a shadow market in satisfying it.
WHY AI ADOPTION IS ESPECIALLY GAMEABLE
Most KPIs measure outputs that are hard to fabricate — units shipped, revenue booked, bugs closed. AI usage metrics typically measure prompt counts or session minutes, which cost nothing to manufacture. The marginal cost of feeding a copilot a fake task is roughly zero, so the equilibrium adoption rate under mandates approaches whatever the quota demands.
THE WELLS FARGO PARALLEL
Wells Fargo's 2016 scandal — 3.5 million fake accounts opened to hit cross-sell targets — was Goodhart's Law at industrial scale. The bank measured accounts-per-customer because it correlated with profitability historically; once it became the target, employees opened accounts customers never asked for. The metric kept rising as the underlying relationship inverted.
THE MANAGEMENT FIX
The literature is consistent: metric-driven mandates work only when the metric is expensive to fake, audited by an independent party, and paired with an outcome measure that catches divergence. AI adoption KPIs typically satisfy none of these. The honest path is measuring downstream productivity — code shipped, tickets resolved, time-to-decision — and letting tool usage emerge as a consequence.