Superalignment
Superalignment is the problem of aligning AI systems much smarter than humans, a term popularised by OpenAI's 2023 Superalignment team.
What is Superalignment?
Current alignment relies on human judgement, which may fail for superhuman systems. One research direction studies whether weak supervisors can train strong models (weak-to-strong generalisation).
How it works
Research uses automated alignment researchers, scalable oversight and interpretability.
← All Artificial Intelligence & Superintelligence definitions · ← Back to the A–Z definitions