AI alignment

The work of making AI systems pursue the goals and values their designers and users intend, and avoid harmful or deceptive behavior.

Alignment covers practical techniques used in today’s assistants — instruction tuning, reinforcement learning from human feedback, written principles and refusal training — as well as research on keeping much more capable future systems under meaningful human control.

Open problems include specifying goals precisely, detecting when a model behaves differently under evaluation, and understanding what happens inside models. Read what AI alignment is.

#Glossary #Safety