Glossary
Guardrails — Glossary
Technical and policy controls that constrain AI inputs and outputs to reduce harmful or out-of-policy behavior.
Full definition
Guardrails include classifiers, allow/deny topic filters, tool permissioning, rate limits, PII redaction, and human review queues. They complement—not replace—model alignment training.
Measure both underblocking (harmful content passes) and overblocking (benign content blocked). Document residual risk using frameworks such as NIST AI RMF.