Rules and constraints built into an AI system to prevent harmful, off-topic, or unintended outputs.
Guardrails can be implemented at the prompt level, through output filtering, or via separate validation models — critical for any AI application used by real users.
