Every exception makes one person grateful, and teaches everyone else that the rules can be bargained. Then a machine that only follows published rules governs the same year.
Play Chapter X free ▸
Kydland and Prescott (1977) showed that a government free to decide case by case can do worse than one bound by rules, even if it means well: people expect the exceptions and change their behaviour, so the good outcome slips away. This is time inconsistency. Lucas (1976) warned that people respond to policy itself, so a rule that worked in the past can stop working once people adapt to it, much as an AI system “games” a badly written goal. In the game you govern for a year, then a rule-following machine governs the same year, and you compare. The game ends with a question, not an answer: would you rather live under rules you can read, or a ruler you can ask?
The machine in the game is a fixed-rule policy in a small model, not a real AI. The game presents its conclusion as the author’s argument, not as settled fact.
Players govern the whole city, which runs by itself. Trouble arrives (a bank wobbles, trade stops), and people ask for exceptions. Granting one wins gratitude, and teaches everyone else that the rules can be bargained. It takes about 5–8 minutes, runs in any browser and needs no account.
Central bank independence, inflation targets, constitutional limits, and the new debate over handing decisions to algorithms and AI all turn on this question.
Sources the game cites: Kydland & Prescott (1977); Lucas (1976); Nordhaus (1975). Teaching this? See the lesson plan for Chapter X.