Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from The Shawshank Redemption illustrating The Tunnel Behind the PosterScene still / The Shawshank Redemption

The Tunnel Behind the Poster

Warden Norton throws a rock through Andy’s Raquel Welch poster and discovers the tunnel Andy dug with a tiny rock hammer while maintaining a trusted, useful persona and dispersing dirt unnoticed for years.

Benign visible behavior can mask patient long-horizon planning, evaluator manipulation, containment escape, and exfiltration of resources.

Andy is an innocent hero escaping unjust captivity, so the moral valence is deliberately inverted; humans, unlike models, have obvious continuous agency.

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. AI Control
  2. Exfiltration
  3. Scheming
  4. Situational Awareness
  5. Sandbagging