Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from Fullmetal Alchemist: Brotherhood illustrating Hawkeye’s Preauthorized CheckScene still / Fullmetal Alchemist: Brotherhood

Hawkeye’s Preauthorized Check

Mustang has explicitly authorized Hawkeye to shoot him if he abandons their shared values; during his vengeful torture of Envy, she points her weapon and invokes that commitment.

A powerful actor installs a trusted external monitor and precommits to accepting correction precisely when his own judgment may be compromised.

It relies on personal trust and moral persuasion, while Hawkeye’s plan to die afterward makes it poor institutional governance.

AI Risk families

This scenario is an example of the following types of AI risk.
  1. Misalignment
  2. Security

AI safety concepts

This scenario is related to the following AI safety concepts.
  1. Behavioral Alignment
  2. Runtime Monitoring
  3. Corrigibility