In plain English
Guardrails are the rules, filters, and constraints placed around an AI system to keep it from producing harmful, off-topic, or policy-violating outputs. They can sit in the prompt, in separate checking systems, or in the surrounding application logic. Good guardrails are layered, because no single check is fully reliable on its own.
Why it matters
Guardrails are how a capable but unpredictable model is made safe enough to put in front of customers, by constraining what it is allowed to say and do.
A worked example
A customer-service assistant configured to decline legal or medical advice and to refuse requests outside its remit is operating within guardrails.
Common confusion
Guardrails are not foolproof. Determined users can sometimes get around them, so they reduce risk rather than remove it.