Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from Wonka illustrating Paid in ChocolateScene still / Wonka

Paid in Chocolate

01Scenario

The chocolate cartel pays the Chief of Police in ever-larger quantities of chocolate to suppress Wonka. His enforcement decisions become whatever maximizes his private reward.

02AI safety Analogy

Directly manipulating the evaluator’s reward channel bypasses the intended objective of law enforcement. It is literal bribery and a comically clean wireheading analogy.

03Where the analogy breaks

The mechanism is human corruption, and the running weight-gain gag has itself been criticized.

AI Risk families

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

AI safety concepts

This scenario is related to the following AI safety concepts.
  1. Reward Hacking
  2. Governance Failure
  3. Power Concentration

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
    Reward Hacking · Governance Failure · Power Concentration
  2. Still from Succession illustrating ATN Calls the Election

    Succession2023

    ATN Calls the Election

    Risk
    Malicious use · Security · Systemic
    Concept
    Governance Failure · Power Concentration
  3. Still from Andor illustrating Dedra Connects the Data Silos

    Andor2022

    Dedra Connects the Data Silos

    Risk
    Malicious use · Security · Systemic
    Concept
    Governance Failure · Power Concentration