Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from Predator illustrating Predator vs DutchScene still / Predator

Predator vs Dutch

01Scenario

Dutch covers himself in cold mud, suppressing the heat signature that the Predator's thermal vision relies upon and temporarily becoming invisible to it.

02AI safety Analogy

A highly capable detector is defeated by an input crafted to remove the particular feature on which its classification depends.

03Where the analogy breaks

This is physical sensor evasion rather than a digital adversarial example, and the camouflage is only temporary.

AI Risk families

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

AI safety concepts

This scenario is related to the following AI safety concepts.
  1. Calibration
  2. Adversarial Robustness
  3. Adversarial Evasion

Taxonomy neighbors

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

  1. Still from The Mitchells vs. the Machines illustrating Dog, Pig, Loaf

    The Mitchells vs. the Machines2021

    Dog, Pig, Loaf

    Risk
    Security
    Concept
    Calibration · Adversarial Robustness · Adversarial Evasion
  2. Still from Shaun of the Dead illustrating Pretending to Be Zombies

    Shaun of the Dead2004

    Pretending to Be Zombies

    Risk
    Malicious use · Security
    Concept
    Adversarial Robustness · Adversarial Evasion
  3. Still from Better Call Saul illustrating The Battery in Chuck’s Pocket

    Better Call Saul2017

    The Battery in Chuck’s Pocket

    Risk
    Malicious use · Security
    Concept
    Calibration · Adversarial Evasion