Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from The Rehearsal illustrating Planting the Trivia AnswersScene still / The Rehearsal

Planting the Trivia Answers

01Scenario

Nathan covertly exposes Kor to the night’s trivia answers through staged encounters. Kor later appears to ace the test without realizing that its answer key leaked into his preparation.

02AI safety Analogy

A benchmark can report excellent performance while measuring contamination rather than the capability it was designed to test.

03Where the analogy breaks

Kor is unaware of the leakage, whereas an advanced model might recognize and exploit it strategically.

AI Risk families

This scenario is an example of this type of AI risk.
  1. Security

AI safety concepts

This scenario is related to the following AI safety concepts.
  1. Deployment Safety
  2. Capability Evals
  3. Data Poisoning
  4. Sandbagging

Taxonomy neighbors

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

  1. Still from Guardians of the Galaxy Vol. 2 illustrating Baby Groot’s Death Button

    Guardians of the Galaxy Vol. 22017

    Baby Groot’s Death Button

    Risk
    Security
    Concept
    Deployment Safety · Capability Evals
  2. Still from Cyberpunk: Edgerunners illustrating Cyberskeleton Field Test

    Cyberpunk: Edgerunners2022

    Cyberskeleton Field Test

    Risk
    Security
    Concept
    Deployment Safety · Capability Evals
  3. Still from Frieren: Beyond Journey’s End illustrating Frieren Suppresses Her Mana

    Frieren: Beyond Journey’s End2023

    Frieren Suppresses Her Mana

    Risk
    Security
    Concept
    Capability Evals · Sandbagging