Safety · Fast-moving · Advanced

AI Alignment

Making an AI system pursue the goals its developers and users actually intend, including implicit expectations they never stated.

What AI Alignment is

Alignment splits into outer alignment — specifying the right objective — and inner alignment — ensuring the trained system actually adopts it rather than a correlated proxy.

How it works

Current practice relies on instruction tuning, preference optimisation, constitutional methods, evaluation against behavioural specifications and interpretability research to inspect internals.

Why it matters

As systems take more autonomous action, the gap between what we asked for and what we meant becomes an operational risk rather than a philosophical one.

Common uses

→Assistant behaviour specification
→Refusal and safety policy
→Agent objective design

Strengths

✓Directly improves everyday usefulness as well as safety

Watch for

✓Whose values is a contested question
✓No verification method for general behaviour

Continue exploring

Reinforcement Learning from Human Feedback Constitutional AI Specification Gaming Reward Model Interpretability