Dernière mise à jour :
Predefined constraints or checks (technical and policy) embedded in AI systems to prevent unsafe or non-compliant behavior at runtime.
Defensive layers - both algorithmic (confidence thresholds, input sanitization, adversarial detectors) and procedural (approval gates, human-in-the-loop) - that enforce policy boundaries. Guardrails limit outputs (e.g., no hate speech), restrict decision domains, and trigger fail-safe actions. Effective guardrail governance requires reviewing and updating constraints as new threats emerge and monitoring bypass attempts.
Exemple concret :
A content-moderation AI has guardrails that block profanity and extremist content. When the NLP filter’s confidence is below 70%, the system routes the post to a human moderator rather than posting automatically - ensuring unsafe content never reaches the feed without oversight.
