Guardrails
The controls placed around an AI system to keep its behaviour within safe, acceptable limits — filtering harmful requests, blocking unsafe outputs, and keeping it on task. Guardrails sit around a model rather than inside it.
A capable AI model will, left to itself, do more or less whatever it is asked, including things it should not: giving dangerous instructions, producing offensive content, or wandering well outside the job it was deployed for. Guardrails are the checks put in place to prevent that. They can inspect what a user asks, filter or block what the model produces, and confine it to approved topics and actions.
The important thing to understand is that guardrails generally wrap around the model rather than living within it. Training methods like reinforcement learning from human feedback shape a model’s own tendencies; guardrails add an extra layer of rules and filters on top, so that even if the model would otherwise misbehave, the surrounding system catches it. Most serious AI products use both.
Guardrails are not foolproof. People probe them for gaps — a practice known as red teaming — and clever phrasing can sometimes slip past them. But they are a core part of deploying AI responsibly, and the phrase turns up whenever a company explains how it keeps its system safe for real users.