Knowledge Maxxing← Back to topic library
AI MaxxingIntermediate75 sec
AI guardrail
An AI guardrail is a check applied before or after a model generates output, to block or adjust responses that violate defined safety or policy rules.
Prerequisite: System prompt
What matters
Guardrails add a layer of control the model's own training might not fully guarantee, catching unwanted outputs like unsafe content or policy violations before they reach a user.
Example
A guardrail might scan a model's draft response for personal data before sending it, and block or redact the reply if any is found.
Common misconception
A guardrail is not part of the model itself; it is typically separate logic wrapped around the model's input or output.
Recall question
Why might a system need a guardrail in addition to the model's own training?
Keywords and related concepts
guardrailsafety checkcontent filteringmodel gateway