Alignment

Alignment is the field and practice of getting AI systems to do what people actually want, including things they did not spell out, and to avoid harmful behaviour. For today's assistants, that means being helpful, honest, and harmless: following instructions sensibly, not inventing facts, refusing clearly dangerous requests, and not manipulating users.

Practical alignment work includes training methods such as RLHF and constitutional AI, written model specifications or constitutions describing intended behaviour, red-teaming to find failures, and evaluations that measure honesty, refusal accuracy and other traits. Interpretability research tries to understand what is happening inside models so problems can be found directly rather than only from outputs.

Alignment is also a longer-term research concern. As AI systems become more capable and act more independently, as agents that use tools and take actions, the gap between what we specified and what we intended matters more. A misread instruction in a chat is an annoyance; the same misreading in an agent with access to email or payments can cause real harm.

Example

A user asks an assistant to "make my essay sound smarter". A poorly aligned system might add impressive but false citations. An aligned one improves the writing and says it can not add sources it has not verified.

Common mistake

Thinking alignment only means refusing harmful requests. Over-refusing safe requests is also an alignment failure, as is sycophancy or quietly doing a different task from the one asked. Alignment is about doing the intended thing well, not just avoiding bad things.