Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from Terminator 2: Judgment Day illustrating The Wolfie TestScene still / Terminator 2: Judgment Day

The Wolfie Test

The T-1000 impersonates John’s foster mother over the phone. Knowing the dog is named Max, the T-800 deliberately calls him Wolfie; the impersonator accepts the false name, trips the test, and reveals that the house is compromised.

A simple challenge-response canary tests whether a convincing voice and identity claim is backed by the right private knowledge. The defense detects the spoof before trusting it or entering the compromised environment.

This is a one-shot authentication tripwire rather than continuous monitoring of an AI system’s internal state.

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. Adversarial Robustness
  2. Runtime Monitoring
  3. Synthetic Fraud