Guardrails

Guardrails is a catch-all term for the safety mechanisms that keep AI from swerving into dangerous territory—like the guardrails on a highway. Training the model to refuse dangerous requests, filters that check inputs and outputs, and limits on what actions an agent can take all count as guardrails.

The term comes up especially often in stories about companies adopting AI. An "in-house chatbot with guardrails" means safeguards have been put in place so it won't leak company secrets or take actions outside its authority. But tighten the guardrails too much and you get "over-refusal,"where even perfectly reasonable questions get rejected—so how tight to make them is a constant point of debate.