Guardrails are checks around an AI model's inputs and outputs, such as blocked topics, format validation, or approval steps, that keep it inside agreed limits.
Category
AI
Also known as
AI guardrails; input filter; output filter; content filter
The ones you control live in the software around the model, on top of whatever safety training the vendor built in. Examples: a rule that any discount above a set threshold needs human sign-off, or that generated copy must pass a brand filter before it is sent.
In practice
Write guardrails for the actions as well as the words. A model that cannot say anything off-brand but can send to the wrong segment is still unguarded where it matters.
Something out of date?