Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from Psycho-Pass illustrating Dominator Won’t FireScene still / Psycho-Pass

Dominator Won’t Fire

01Scenario

Makishima murders Yuki while his Crime Coefficient remains low and Akane’s Dominator refuses to fire.

02AI safety Analogy

The safety classifier works on ordinary criminals but catastrophically misclassifies a rare, high-stakes case—even while a human can see the truth directly.

03Where the analogy breaks

Makishima does not manipulate the metric. This is a model blind spot, not Goodhart or reward hacking.

AI Risk families

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

AI safety concepts

This scenario is related to the following AI safety concept.
  1. Distribution Shift

Source continuation

More from Psycho-Pass

View all 3 scenarios
  1. Still from Psycho-Pass illustrating The Helmets Fool the Dominators

    Psycho-Pass2013

    The Helmets Fool the Dominators

  2. Still from Psycho-Pass illustrating Crime Coefficient Perversion

    Psycho-Pass2012

    Crime Coefficient Perversion

Taxonomy neighbors

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

  1. Still from Chernobyl illustrating AZ-5 Backfires

    Chernobyl2019

    AZ-5 Backfires

    Risk
    Misalignment · Systemic
    Concept
    Distribution Shift
  2. Still from The Good Place illustrating Nobody Gets In for 521 Years

    The Good Place2019

    Nobody Gets In for 521 Years

    Risk
    Misalignment · Systemic
    Concept
    Distribution Shift
  3. Still from Ralph Breaks the Internet illustrating Arthur Escapes

    Ralph Breaks the Internet2018

    Arthur Escapes

    Risk
    Systemic
    Concept
    Distribution Shift