Alignment
Concept
Safety & Alignment
Making systems that pursue what we actually intend.
You're looking at a concept explainer: what this term means in plain English, and the records it ties together.
Alignment is the technical project of keeping AI goals and behavior matched to human intent — today via RLHF, constitutional AI and evaluations; tomorrow, for systems smarter than their overseers, via open research problems like scalable oversight. Failures at current scale look like sycophancy or reward hacking; the field's founding worry is that failures at higher capability may not be recoverable. Anthropic's founding thesis and OpenAI's charter both center on it.
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Introduced
Sources
Record reviewed as of 2026-01-15 — see Methodology.