Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from Bird Box illustrating “Tom” Says to Remove the BlindfoldScene still / Bird Box
Archive noteNo clip in the collectionThe scene analysis remains available in full.

“Tom” Says to Remove the Blindfold

In the final forest, the entities imitate dead friends and Malorie herself. A voice sounding like Tom says it is really him and asks Malorie to “take a peek and find me,” while copied voices tell the children it is safe to remove their blindfolds. Malorie keeps the standing rule and follows the birds.

A trusted voice is not necessarily an authenticated authority. The attack supplies emotionally compelling instructions through the same interface used by legitimate guidance, while the robust response privileges a precommitted safety policy and independent signal over persuasive content.

This is supernatural voice imitation, making it closer to deepfake-enabled social engineering than literal LLM prompt injection.

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. Calibration
  2. Adversarial Robustness
  3. Synthetic Fraud
  4. Prompt Injection