Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from Altered Carbon illustrating Bancroft’s Stale Backup

Bancroft’s Stale Backup

01Scenario

Laurens Bancroft is restored from a remote backup after his sleeve is killed, but the 48-hour backup gap erases the memory of what happened and who killed him.

02AI safety Analogy

Checkpointing preserves capability while discarding the very evidence needed for diagnosis and accountability.

03Where the analogy breaks

A human memory backup is not an AI model checkpoint, but the observability failure is closely analogous.

AI Risk families

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

AI safety concepts

This scenario is related to the following AI safety concepts.
  1. Contestability
  2. Runtime Monitoring
  3. Power Concentration
  4. Interpretability

Source continuation

More from Altered Carbon

View all 3 scenarios
  1. Still from Altered Carbon illustrating Dimi’s Double Sleeve

    Altered Carbon2018

    Dimi’s Double Sleeve

  2. Still from Altered Carbon illustrating Lizzie Trains in VR

    Altered Carbon2018

    Lizzie Trains in VR

Taxonomy neighbors

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

  1. Still from True Detective illustrating Tuttle Captures Oversight

    True Detective2014

    Tuttle Captures Oversight

    Risk
    Security · Systemic
    Concept
    Contestability · Power Concentration
  2. Still from Lost illustrating Push the Button

    Lost2006

    Push the Button

    Risk
    Security
    Concept
    Runtime Monitoring · Interpretability
  3. Still from Anatomy of a Fall illustrating The Tape Is Not the Marriage

    Anatomy of a Fall2023

    The Tape Is Not the Marriage

    Risk
    Systemic
    Concept
    Contestability · Interpretability