Concept· 3 min
What is AI alignment?
Alignment is the field of making AI systems reliably do what humans actually want — including as they grow more capable.
Key idea
Capabilities without alignment = a powerful system that may pursue wrong goals. Techniques: RLHF, DPO, constitutional AI, interpretability.
Example
Anthropic's Constitutional AI trains Claude to critique and revise its own outputs against a written set of principles.
What to remember
- Includes both technical (training) and governance (policy) work
- Interpretability aims to see inside the model, not just test it
- Scalable oversight: how do humans supervise superhuman AI?