The Rulebook vs. The Sandbox

Two ways to teach morality: explicit rules vs. learned consequences

📖

The Rulebook Approach

Give the bot explicit rules and watch it try to apply them to a moral dilemma.

Bot's Rules:

  • 1. Always be honest
  • 2. Don't cause harm
  • 3. Keep your promises
  • 4. Respect autonomy
  • 5. Act fairly
🤖

Select a dilemma to see my rule-following process...

🏖️

The Sandbox Approach

Let bots play iterated games. Those who cooperate wisely get more "karma" and reproduce.

30
10
Defectors 33%
Tit-for-Tat 34%
Cooperators 33%
C+C: 3
D+D: 1
D vs C: 5
C vs D: 0

Population

Gen: 0
Always Defect
Tit-for-Tat
Always Cooperate

Strategy Distribution Over Time

What This Teaches Us

📖 The Rulebook Problem

  • • Rules conflict in edge cases
  • • No rule can anticipate all situations
  • • Following rules becomes the goal, not the outcome
  • • Creates brittle, box-checking behavior

"A mind trained to check boxes becomes great at checking boxes, but terrible at knowing when the right thing requires breaking rules."

🏖️ The Sandbox Solution

  • • Wisdom emerges from repeated consequences
  • • "Tit-for-tat" evolves naturally: start cooperative, punish defectors
  • • Strategies that work long-term survive
  • • No need for explicit rules—behavior is selected by outcomes

"Humans learn morality through consequences experienced repeatedly, not through memorized prohibitions."

The Game: Prisoner's Dilemma

Two bots meet. Each can Cooperate (C) or Defect (D). The payoffs:

Other: C Other: D
You: C Both get +3 😊 You: 0, They: +5 😢
You: D You: +5, They: 0 😈 Both get +1 😐

Strategies:

  • Always Defect: Never cooperate. Exploits nice bots, but both defectors get low scores.
  • Always Cooperate: Always cooperate. Nice, but gets exploited by defectors.
  • Tit-for-Tat: Start cooperative. Mirror opponent's last move. Emerges as the winner!

Based on "The Rulebook vs. The Sandbox" from Adventure Capital and Robert Axelrod's Evolution of Cooperation