Goodhart's Law at Amazon: How "Toxenmaxxing" Proves the Surveillance Paradox

## The Incident

Amazon created an AI usage leaderboard. Employees assigned AI agents to meaningless tasks to climb the rankings. The behavior became so common that Amazon coined a term for it: "toxenmaxxing." Then they sent a memo telling employees to stop.

```mermaid
graph TD
    A[Company wants: More AI adoption] --> B[Creates: AI usage leaderboard]
    B --> C[Employees: Assign AI to pointless tasks]
    C --> D[Metric: AI usage goes up]
    D --> E[Company realizes: Productivity NOT up]
    E --> F[Sends memo: "Don't use AI just to use AI"]
    F --> G[Employees: Find next metric to game]
```

## Goodhart's Law, Formally

"When a measure becomes a target, it ceases to be a good measure."

Every surveillance technology in workplace history — from time clocks in 1890 to AI leaderboards in 2026 — has been gamed within months of deployment. The pattern is so consistent it should be taught as a management axiom.

## The Alternative

Measure outcomes, not processes. "Did the project ship?" — hard to game. "How many AI queries did you run?" — trivially gameable. The difference is trust: outcome measurement trusts employees to figure out how. Process measurement assumes they will cheat if not watched. They do cheat. Not because they are bad employees, but because the system was designed to be cheated.

Goodhart's Law at Amazon: How "Toxenmaxxing" Proves the Surveillance Paradox

// COMMENTS

ON THIS PAGE