AI Glossary · Definition
What is Guardrails?
Guardrails are rules and technical controls that keep an AI system within safe, approved behaviour, such as topics it may discuss, data it may use and actions it may take.
By DAIDU EditorialUpdated
Guardrails, explained
Guardrails can be written into instructions ('never quote prices that are not in the price list'), enforced by software (blocking certain data from leaving a system), or built into workflows (requiring approval before sending). They reduce the risk of harmful, off-brand or incorrect outputs, and of agents taking actions they should not. Guardrails should be tested with deliberately tricky inputs, not just normal ones.
Example
A WhatsApp assistant is set to decline medical questions and hand them to clinic staff immediately.