Alignment is the work of making AI systems do what people actually intend while respecting rules and values.
Example
A model should answer helpfully but give no instructions for dangerous substances, even when someone asks cleverly.
For pros
Methods include RLHF, written specifications for model behaviour, red teaming and interpretability research. How to reliably test alignment in very capable systems is an open question. The term is sometimes used broadly for a model’s “safety and behaviour”.
