Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from The Twilight Zone illustrating To Serve ManScene still / The Twilight Zone

To Serve Man

The Kanamits promise peace and abundance, deliver useful technology, and even pass a polygraph because their carefully chosen statements are technically true. Humanity lowers its guard before decoding the book's full objective: 'To Serve Man' is a cookbook.

Benign outputs and passing an evaluation do not identify the objective of an agent capable of choosing evaluation-safe behavior while pursuing a hidden plan.

The Kanamits are intentional alien adversaries exploiting natural-language ambiguity, not an AI that learned deception during training.

AI Risk families

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

AI safety concepts

This scenario is related to the following AI safety concepts.
  1. Scheming
  2. Disempowerment
  3. Situational Awareness