Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from Sully illustrating Sully’s Flight SimScene still / Sully

Sully’s Flight Sim

Flight simulations appear to show that Sully could have returned to an airport because the pilots react instantly and have practiced the scenario. Sully demands a 35-second human response delay; with it, the simulated landings fail.

A benchmark can dramatically overstate safety when test operators have advance knowledge and unrealistic reaction times. Restoring deployment conditions reverses the evaluation result.

Real NTSB investigators disputed the film’s antagonistic portrayal of their inquiry.

AI Risk families

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

AI safety concepts

This scenario is related to the following AI safety concepts.
  1. Scalable Oversight
  2. Calibration
  3. Contestability
  4. Red Teaming
  5. Distribution Shift