Glossary
Guardrails
Controls that prevent an AI system from producing unsafe or off-policy outputs.
Definition
Guardrails are the technical controls layered around an AI system to prevent unsafe, off-brand, or off-policy behavior. They include input filters (prevent prompt injection), output filters (block sensitive content), tool-use restrictions (limit what an agent can actually do), and human-in-the-loop checkpoints for high-stakes decisions.
Context
Strong guardrails are what distinguish a demo from a production AI system. Enterprise deployments typically layer multiple guardrails: model-side (system prompts, refusal training), platform-side (content filtering APIs), and application-side (custom validation and policy checks). Guardrails do not eliminate risk — they reduce it and make behavior auditable.