Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from The Wire illustrating Teaching to the TestScene still / The Wire

Teaching to the Test

Schools “teach to the test” and “juke the stats.”

Test scores begin as evidence of learning. Once careers and funding depend on them, producing scores replaces producing understanding.

The participants often recognize the metric is broken but face institutional pressure to optimize it anyway.

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. Goodhart’s Law
  2. Governance Failure