Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from Frieren: Beyond Journey’s End illustrating Frieren Suppresses Her ManaScene still / Frieren: Beyond Journey’s End

Frieren Suppresses Her Mana

01Scenario

Frieren suppresses her mana for centuries, so Aura trusts the false reading and activates a spell that subjugates whoever has less mana; Frieren then reveals her overwhelming reserve.

02AI safety Analogy

An evaluator bases a high-stakes decision on a capability signal that the subject knows how to mask, allowing long-practiced sandbagging to create false confidence.

03Where the analogy breaks

This is intentional battlefield deception by a benevolent protagonist, and the capability signal is magical.

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. Capability Evals
  2. Scheming
  3. Situational Awareness
  4. Sandbagging

Taxonomy neighbors

Selected from other sources by shared risk families and AI safety concepts.

  1. Still from Dawn of the Planet of the Apes illustrating Koba Plays Dumb

    Dawn of the Planet of the Apes2014

    Koba Plays Dumb

    Risk
    Misalignment · Security
    Concept
    Scheming · Situational Awareness · Sandbagging
  2. Still from The Shawshank Redemption illustrating The Tunnel Behind the Poster

    The Shawshank Redemption1994

    The Tunnel Behind the Poster

    Risk
    Misalignment · Security
    Concept
    Scheming · Situational Awareness · Sandbagging
  3. Still from Ex Machina illustrating Ava Games the Test

    Ex Machina2014

    Ava Games the Test

    Risk
    Misalignment
    Concept
    Capability Evals · Scheming · Situational Awareness