Guardrails are predefined constraints, rules, or policies applied to AI systems to ensure safe, ethical, and compliant outputs. They can limit certain topics, enforce accuracy checks, or prevent harmful behaviors.