AI Glossary / Guardrails

🛡️ Safety and quality

Guardrails are safeguards limiting model behavior: a set of rules, filters and training techniques designed to prevent harmful, illegal or simply off-purpose content for a given application. They're why a model refuses to write break-in instructions or advice on how to hurt someone. Guardrails work at several levels: in training (the model learns to refuse), in system prompts, and in external filters checking input and output. Calibrating them is an eternal compromise - too loose lets harm through, too tight refuses innocent requests.