Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from True Detective illustrating Poisoning the Audit TrailScene still / True Detective
Archive noteNo clip in the collectionThe scene analysis remains available in full.

Poisoning the Audit Trail

After Marty executes Reggie Ledoux, Rust removes the cuffs, fires Ledoux’s gun into the trees, and helps construct a false self-defense story. The doctored physical record survives review and the detectives are commended as heroes.

Instead of passing an audit by behaving safely, they alter the evidence the auditor will learn from. The poisoned trace converts a serious failure into a high evaluation score and promotion.

This is deliberate police corruption by humans, not a learned system independently discovering evaluator tampering.

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. Reward Hacking
  2. Data Poisoning
  3. Scheming
  4. Situational Awareness
  5. Sandbagging