Inner Alignment

Model behaves well during training but not deployment.

Learned subsystem that optimizes its own objective.

Maintaining alignment under new conditions.

Willingness of system to accept correction or shutdown.

Maximizing reward without fulfilling real goal.

Tendency for agents to pursue resources regardless of final goal.

Model optimizes objectives misaligned with human values.

Correctly specifying goals.