Alignment
AI Alignment
ActiveUpdated: 28 August 2026Published
Key
innovation
Ensuring that AI systems' goals and behaviors match human intentions and values.
Category
Alignment
Abstraction level
Paradigm
Operation level
Architecture blockModel
Use cases
Training AI assistants (helpful/harmless/honest)Frontier-model safetyOversight and control of agentsAI policy and regulation
How it works
Techniques include reinforcement learning from human feedback (RLHF) and from AI feedback (RLAIF), constitutional AI, scalable oversight, interpretability, red-teaming, and capability/threat evaluations (e.g. work by METR and Palisade Research).
Problem solved
It is hard to precisely encode human intentions and values into an optimization objective; misaligned, capable systems may act in undesirable or unsafe ways.