Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from Twin Peaks illustrating Cooper Cannot Retrieve the AnswerScene still / Twin Peaks

Cooper Cannot Retrieve the Answer

In Cooper's Red Room dream, Laura says that she feels as though she knows who killed her and whispers the answer in his ear. Cooper wakes convinced that he knows the killer, yet the knowledge is inaccessible until later clues make the latent answer recoverable.

A system can contain decision-relevant knowledge in a latent representation while failing to expose it in a reliable, immediately actionable form.

The source is supernatural dream logic rather than a trained model, and the whispered answer ultimately is correct.

AI Risk families

This scenario is an example of this type of AI risk.
  1. Accidents

AI safety concepts

This scenario is related to the following AI safety concepts.
  1. Calibration
  2. Knowledge Elicitation
  3. Interpretability