AI Control
AI control is a research approach that aims to make AI deployments safe even if the models might be misaligned, using monitoring and restrictions that the models cannot subvert.
What is AI Control?
Rather than relying solely on alignment, control protocols assume the worst and test whether safeguards hold against a model actively trying to cause harm.
How it works
Trusted weaker models monitor stronger untrusted ones, suspicious actions are audited and permissions are limited.
← All Artificial Intelligence & Superintelligence definitions · ← Back to the A–Z definitions