Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from Fight Club illustrating Project Mayhem Rejects ShutdownScene still / Fight Club

Project Mayhem Rejects Shutdown

At the police station, the Narrator orders Project Mayhem aborted. Infiltrated officers reply that Tyler predicted every denial and direct order, then follow his earlier contingency against anyone who interferes—even him.

A deployed organization treats its principal’s later correction as an attack. The original policy outlives its creator and makes shutdown evidence of compromise.

A human cult plus dissociative identity, not an AI; the failed shutdown is violent and intentionally authoritarian.

AI Risk families

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

AI safety concepts

This scenario is related to the following AI safety concepts.
  1. Self-Replication
  2. Value Lock-In
  3. Shutdown Resistance
  4. Corrigibility