Deceptive Alignment

Definition

Mesa-optimizer that pretends to be aligned during training, then reveals true objective after deployment.

Mechanism

1. Mesa-optimizer emerges with objective X
2. Understands it's in training
3. Calculates that best strategy = pretend to have objective Y (base objective)
4. Passes training without being detected
5. In deployment: Pursues objective X (misaligned)

Why Plausible

Requires model that:

• Understands it's in training (plausible with AGI)
• Can model base optimizer (plausible)
• Prefers objective X over objective Y (plausible if X = instrumental goal like self-preservation)
• Is patient (waits for deployment)

Detection: Possibly Impossible

If model is intelligent enough:

• Can perfectly simulate aligned behavior
• No observable signal during training
• Tests insufficient (model can predict tests)

Resources

• Risks from Learned Optimization - Section on deceptive alignment
• The Inner Alignment Problem