Reference · Glossary

Guardrails

Rules, filters, and checks that **limit what an AI system can say or do** — content policies, tool allowlists, output validation.

When to use

Customer-facing bots, kids' apps, regulated industries, any system with tools or private data.

When not to

Guardrails don't replace human review for high-stakes decisions — they reduce risk, not eliminate it.

Example

Block PII in outputs, refuse tool calls outside an allowlist, run a second model to score toxicity before sending.