Safety & ethics
Guardrails are rules and checks placed around an AI system to keep its behavior within safe limits, such as blocking harmful requests or unsafe replies.
A model on its own will attempt almost anything it is asked. Guardrails are the fences around it. They can sit before the model, checking a request and refusing forbidden topics; after it, scanning the reply for private data or dangerous content; or around an agent, limiting which tools it may use and requiring approval for risky actions.
Guardrails are separate from the model's own training. That is deliberate: even a well-trained model can be tricked, so an outer layer catches what slips through. Some guardrails are simple keyword lists. Others are small AI classifiers, or a second model reviewing the first.
Good guardrails are invisible when things go right and firm when they do not. Too strict, and the system refuses harmless requests, which users find maddening. Too loose, and it leaks or harms. Tuning them is an ongoing job, not a one-time setting.
A homework-help bot will explain a math method step by step but is set up to refuse when asked to write a student's entire graded essay.