Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from Pantheon illustrating Maddie’s Simulated WorldsScene still / Pantheon
Archive noteNo clip in the collectionThe scene analysis remains available in full.

Maddie’s Simulated Worlds

In deep time, Maddie runs vast numbers of simulated worlds and conscious lives to reconstruct Caspian and find the sequence of events she wants.

At sufficient scale, experimentation can create enormous populations of morally relevant subjects. Oversight of one operator no longer captures the welfare cost hidden inside the search process.

The finale deliberately blurs simulation, reality, and consent; counting and comparing the simulated lives is philosophically uncertain.

AI Risk families

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

AI safety concepts

This scenario is related to the following AI safety concepts.
  1. Value Alignment
  2. Scalable Oversight
  3. AI Welfare